enterprise ai teams

Abstract: Federated Learning (FL) enables collaborative model training without exchanging raw data, but conventional FL provides limited runtime governance, evidence verification, fairness assessment, explainability monitoring, and auditability. This paper proposes Governance-Aware Federated Learning (GAFL), which combines independent governance-evidence verification, temporal trust, fairness and explanation assessment, anomaly detection, policy enforcement, normalized aggregation, and tamper-evident audit logging. Client-reported governance values are authenticated, bound to their submitted model updates, and independently recomputed before aggregation. A cross-silo prototype with 10 participants was evaluated on the Adult Income dataset under honest, benign non-IID, biased-data, malicious-update, non-compliant, falsified-evidence, and mixed-threat scenarios. Four methods were compared across five random seeds: FedAvg, FedAvg with anomaly filtering, GAFL without independent verification, and complete GAFL. Complete GAFL intervened on all affected participants at the final round across the evaluated risk scenarios, including identifying 100% of falsified-evidence participants, whereas GAFL without verification identified none of the falsified-evidence participants. Final macro-F1 ranged from 0.7520 to 0.7733 across GAFL scenarios. Complete GAFL required 173.67 plus or minus 26.94 ms of governance processing per round, compared with 3.82 plus or minus 0.79 ms for FedAvg. A strict demographic-parity threshold also caused frequent interventions for honest participants, demonstrating a significant fairness-policy and utility trade-off.

Abstract: Large organizations consume large language models (LLMs) from several vendors at once, each with its own API dialect, credential model, streaming protocol and pricing. Integrated directly, this heterogeneity yields credential sprawl, format lock-in, unbounded spend and inconsistent governance. We present the design and production operation of the Cosmos LLM Gateway, the governed entry point for enterprise LLM traffic at PayPal. Built on Envoy, it splits work by latency profile: I/O-bound control decisions (authentication, credential brokering, policy and budget evaluation) run in an out-of-process external-processing service, while wire-format translation, token metering and prompt-cache shaping run in-proxy as a Rust WebAssembly filter, so that, by default, streamed response body chunks never wait on an out-of-process call. Models, providers and prices are largely validated data rather than code. Governance is declarative policy evaluated by Open Policy Agent and managed through a control plane with role-scoped authoring, approvals, pilots, shadow-mode rollout, usage notifications and versioned policy bundles. For use cases that require it, a data-loss-prevention layer can block, redact or pseudonymize detected sensitive content before it reaches a provider. Spend scopes are views over per-model usage counters, so a new budget applies to spend already recorded in its window without migration. As of September 2026, the gateway serves more than 12,000 developers and hundreds of business use cases, handling roughly 200 billion tokens per day across more than 100 models at 99.999% availability to business. In a benchmark of the core request path against a synthetic upstream, it adds a median of 4 ms and a p99 of 11 ms of overhead at more than 2,200 requests per second on two 4-vCPU instances. We report the architecture, policy model, evaluation and lessons learned.

Abstract: Long-horizon agents continuously encounter reusable procedural knowledge, yet existing systems typically rely either on model adaptation or external memory, trading persistent learning for rapid reuse. This limitation is especially important for Web agents that repeatedly execute related multi-step workflows. We propose HaMo, a Harness-Model collaborative framework that combines fast non-parametric experience accumulation with slow parametric adaptation. HaMo distills trajectories into strategy-level experience cards for immediate retrieval, while an experience-conditioned continual learning objective trains the model to interpret, ground, and execute retrieved strategies across changing contexts. A retrieval-coverage-based selection mechanism further identifies compact and complementary trajectories for parameter updates. Experiments show that HaMo outperforms harness-only and model-only adaptation under within-domain and cross-domain shifts, improving both retention and adaptation. These results highlight the importance of jointly preserving procedural experience and internalizing the competence to transfer and execute it across tasks.

Abstract: Continual learning is the ability of LLM agents to improve over time by acquiring, refining, and applying knowledge from experience while retaining useful prior knowledge. It is essential for agents learning over long time horizons, yet current evaluations largely assess individual task performance rather than improvement through accumulated experience. We introduce Continual Learning Exam (CLE), a large-scale benchmark developed by 50+ domain experts to measure experience-driven improvement in realistic professional workflows. CLE comprises 120 human-authored tasks, each organised as a five-level ladder, yielding 600 levels spanning 89 professional roles across 12 occupational domains. Agents must pass each level to unlock the next, with later levels building on experience accumulated during earlier assignments. An evaluation agent assesses each submission against human-authored rubrics. We compare a single-submission setting with one that provides diagnostic feedback after failure and allows one revision. Across 17 harness-model configurations using frontier LLMs, feedback and revision yield a median completion gain of 5.3 percentage points. Yet the best-performing system clears only 47.7% of levels and completes all five levels in just 23.3% of tasks, highlighting the difficulty of sustained learning from experience for existing frontier LLMs. As such, the dataset can also be used to post-train LLMs to strengthen their continual learning capabilities.

Abstract: Federated continual learning trains a model on a sequence of tasks across several clients without sharing their data while retaining knowledge of earlier tasks. Sparse mixture-of-experts models support this setting by adapting selected parameters on a frozen backbone. However, clients learn different expert content as tasks evolve, making it difficult to share knowledge across clients without disrupting task-specific and client-specific specialization. Matching tile coordinates alone does not establish whether their learned content is compatible. We present FedSEAM, a federated continual learning framework that organizes task knowledge into sparse experts and selectively shares their content across tasks and clients. On each client, task-aware tile selection allocates capacity to a new task, while a subspace-based compatibility test determines which existing expert tiles to reuse and where to introduce fresh capacity. On the server, Subspace Expert Agreement Merging (SEAM) compares client updates as subspaces and moves each client towards consensus only along agreed-upon directions, leaving the rest of its update unchanged. We identify two remaining sources of forgetting, routing drift and merge erasure, and address them with three protections, one of which exactly preserves tile components in a selected prior-overlap subspace at every merge. We evaluate FedSEAM on TRACE using Qwen3-4B and Llama-3.2-3B, with component-wise ablations and studies of data size to assess its balance between learning new tasks and retaining prior knowledge.

Abstract: AI agents capable of autonomous decision-making introduce significant challenges related to safety, alignment, accountability, and human oversight. Existing Human-in-the-Loop (HITL) and responsible AI approaches provide important foundations, but many remain fragmented across system design, governance, and operational deployment. This paper presents a Three-Pillar (3P) model and framework for developing safe and trustworthy AI agents through transparency, accountability, and trustworthiness. The framework defines an evolutionary path in which increasing agent autonomy is introduced through progressive validation, structured human oversight, and auditable operational controls. We further describe how the model supports agent lifecycle monitoring, decision traceability, escalation management, and human-agent collaboration within enterprise agentic environments. By integrating oversight principles with operational mechanisms, the proposed framework provides a practical foundation for deploying AI agents that remain aligned with human intent while enabling controlled progression toward higher autonomy.

Abstract: Computer-use agents are limited by the interfaces humans give them. An agent may type a spreadsheet cell by cell when a single line of code would do, or print all two thousand lines of a file through a shell to find one function. Both the interfaces and the knowledge of when to use each are built by hand, which cannot keep pace with new applications, workflows, and models. We argue that interface design should itself be an agent capability. We introduce Agent S4, an autonomous computer-use scientist that builds Evolving Agent Interfaces from task experience. A meta-agent examines task agents' trajectories on synthetic training tasks, without success labels, and constructs and revises interfaces and usage knowledge in a repository. Population-based evolution then selects which repositories to develop further by their performance on validation tasks the meta-agent never sees, favoring changes that generalize beyond the seen tasks. On OSWorld V2, Agent S4, with Qwen3.8-27B and the evolved repository of interfaces, sets a new state of the art in open-source computer-use agents, reaching 21.3% binary score, 3.8 times the same model with fixed interfaces. This is above both GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%), at about 4.3% and 1.5% of their respective costs. On ScienceBoard, the evolved interfaces improve the success rate from 41.4% to 58.6% with the same model. These results suggest that useful interfaces can be evolved from experience, making autonomous interface design a source of improvement complementary to advances in models and harnesses. We open-source our code at https://github.com/simular-ai/agent-s.

Abstract: Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems further expand these capabilities through specialization and collaboration, but distributed authority, heterogeneous roles, and opaque or potentially compromised participants make safeguarding even more challenging. Existing defenses address important individual threats but provide fragmented coverage across local execution, system-level interactions, and enforcement mechanisms. We introduce MAS-Guard, a unified framework that closes this gap for safeguarding single- and multi-agent AI applications containing both instrumentable and opaque agents. Within instrumentable agents, the proposed local Agent Guardrail (A-Guard) layer applies task-specific experts to assess agentic events. At the system level, developer-owned Agent Charters and a Control Plane combine deterministic structural checks of routes, agent capabilities, and message formats with semantic assessments of roles and message content. These complementary layers enable enforcement through a shared set of interventions: allow, reflection, confirmation, and block. Evaluated across six benchmarks covering diverse single- and multi-agent scenarios with three agent LLMs from different providers, MAS-Guard reduces the overall attack success rate by 71.5-86.3% relative to unguarded agents, while meeting the utility reference criterion on benign tasks.

Abstract: Large language model (LLM) agents increasingly rely on persistent memory to support long-term reasoning and personalized interactions. However, information collected from external sources can introduce security risks when untrusted claims (i.e., atomic statements) are later treated as authorized preferences, policies, or permissions. Existing defenses typically rely on content-based filtering or coarse-grained provenance tracking, which either fail to prevent unauthorized memory use or block legitimate information, sacrificing memory utility. To address these limitations, we propose SAFE-Mem, a fine-grained memory governance framework that improves security while preserving legitimate memory use. First, SAFE-Mem annotates each claim with three governance labels (i.e., kind, authority, and scope) to capture its provenance and permitted use. Second, governance-aware retrieval matches query requirements with claim permissions to block unauthorized influence while retaining legitimate information. Third, a governance-enabled memory graph connects source sentences, claims, and entities to support provenance tracking and efficient retrieval. Across four LLM backbones on MPBench and LoCoMo, SAFE-Mem attains the lowest attack success among defenses that retain legitimate memory, at least 77% below the strongest of them, and the highest answer accuracy of all evaluated defenses, improving it by up to 15%. The code is available at https://github.com/LVenum/SAFE-Mem.

Abstract: Artificial intelligence (AI) for Science is often described as a progression from human-directed assistance toward increasingly autonomous AI scientists. Human-AI collaboration offers a further mode in which human and AI contributions substantively shape scientific reasoning, judgments, and the course of inquiry. Existing research addresses work allocation, human-AI coordination, and system evaluation, but how these aspects interact to shape collaborative scientific systems remains insufficiently articulated. We develop the 9A Framework for Human-AI Collaborative Science (HACS) to connect these considerations across three layers. The organization layer specifies what scientific activities are required (Activity), who performs them (Actor), and who holds authority over consequential decisions (Authority). The coordination layer addresses how participants access relevant research information (Awareness), how scientific work advances through iterative proposal, execution, examination, revision, and resolution (Accountability), and when collaborative arrangements should be adjusted (Adaptability). The evaluation layer establishes what constitutes effective system performance (Achievement), how comparisons can be designed to identify the effects of organizational and coordination choices (Attribution), and what evidence supports the resulting conclusions (Assurance). Together, these layers frame HACS as a system-design problem, providing a basis for designing collaborative scientific workflows and for empirically comparing their effectiveness.

Abstract: Fine-tuning large language models (LLMs) on proprietary records enables domain adaptation but introduces two challenges: (i) privacy leakage of memorized private data, such as personally identifiable information (PII), and (ii) balancing private data removal with utility preservation. Existing defenses, including data sanitization and differential privacy, often degrade performance, while privacy-neuron editing lacks a PII-free reference to distinguish identity-specific neurons from those supporting useful text. We propose JEPA-Prune, a post-hoc neuron pruning framework guided by a model trained exclusively on PII-free views. Alongside the fine-tuned private model, JEPA-Prune, trains a public model on surrogate and masked views of the same records using a joint-embedding predictive architecture (JEPA), capturing domain knowledge without identifiers. A cross-model representation probe quantifies deviations between private representations of leaked records and this PII-free reference. Integrated gradients and a scale-invariant specificity score then identify PII-responsive neurons, which are pruned without retraining. Experiments show that pruning just 1.6% of neurons reduces PII inference accuracy by up to 7.4x and PII extraction recall by up to 13.9x, while keeping unseen-text perplexity within 16% of that of the unpruned model for most LLMs. In contrast, privacy-neuron editing increases perplexity by at least 26x.

Abstract: Retrieval-Augmented Generation (RAG) enhances the capabilities of large language models (LLMs) by querying external structured knowledge. However, it can also introduce privacy risks by leaking sensitive information from the retrieval datastore. We propose a simple method to preserve datastore privacy in RAG systems via random projection. By applying the same projection to both datastore embeddings and query embeddings, our method provably preserves semantic similarity between queries and retrieved items while substantially mitigating data extraction attacks. Across multiple RAG architectures and datasets, we show that this lightweight approach achieves superior retrieval and generation performance compared to prior methods which leverage formal differential privacy (DP) guarantees, while exhibiting comparable empirical privacy under strong attack models. Our results for the first time suggest that random projection can serve as a competitive and practical baseline for privacy-preserving RAG systems.

Abstract: Large language model agents can improve their capabilities by evolving reusable natural-language skills from execution experience. In real-world deployments, this experience is distributed across data owners who cannot pool their raw data, which motivates federated skill evolution. The natural first design lets each client evolve the whole skill document and upload the result for server-side aggregation. That design is poorly matched to skill evolution for three reasons. An evolved document can still carry client-specific content distilled from private traces, repeated transmission of a nearly unchanged document spends communication on knowledge that the population already holds and the contribution of a single client is buried in a document that the server cannot meaningfully attribute. We propose FedSkillGit, a federated skill evolution framework that replaces full-document updates with commits. Each participating client submits a bounded set of patches over named rules together with its local experience, which summarizes what the round taught it, and the pair passes a two-stage privacy filter before it leaves the client. The server deduplicates the submitted patches and the local experience of all participants, updates a global experience document that only the optimizers read, and merges the surviving patches into a candidate skill. The candidate is distributed to the clients, which score it on their own validation sets and report the results, and it is published only when the reported scores improve on the current version. Experiments on three datasets and three language model families show that FedSkillGit outperforms state-of-the-art baselines by up to 22% and approaches centralized evolution without sharing raw client data.

Abstract: We present a graph-based framework for Composite AI systems that integrate multiple AI techniques to address complex problems. We describe an architecture using specialized control nodes to orchestrate diverse models including planning, machine learning, constraint programming, and large language models. The system employs directed acyclic graphs and behavior trees to manage model relationships and execution sequencing. Our methodology encompasses data connectors, computational services, and a containerized deployment framework supporting hierarchical, sequential, and concurrent model composition patterns. Through case studies, we demonstrate how this architecture enables the creation of composite systems that leverage the strengths of multiple AI techniques within a unified operational schema, overcoming the limitations of single-model approaches.

Abstract: Large language model (LLM) routers are increasingly deployed to optimize quality, latency, and cost through dynamic model selection and request rewriting. However, a compromised router can cause ineligible selection on the control plane, unendorsed rewriting or unfaithful relay on the data plane, undermining LLM routing integrity. We present VeriRouter, a verifiable LLM routing framework. Under an endorsed public context binding a private semantic index and rewrite banks, VeriRouter combines ZK proofs with a provider-issued serving certificate to establish integrity across both planes while keeping the router's private routing assets hidden from clients. For model selection, we design an efficient ZK proof for semantic search. For rewriting, local checks enforce message preservation and additions from committed banks. On real-world workloads, VeriRouter's model-selection and rewrite proofs average 21.92~s and 4.24~s, yielding 1,307$\times$ and 101$\times$ speedups over the fastest completing baselines. Relative to serving without verification, VeriRouter adds only 4.08 s under per-turn selection and 0.11 s under session-pinned selection.

Abstract: LLM agents' capabilities depend not only on their underlying models but also on their execution harnesses. Recursive self-improvement (RSI) offers a promising approach to automatically evolve these harnesses. However, deploying RSI in open-world settings still faces two critical challenges. First, open-world tasks are typically unseen during the harness design phase and novel at test time, leaving no ground-truth verifiers to reliably guide online harness edits. Second, the RSI process remains highly task-specific and fails to generalize to novel scenarios. We introduce Harness^2, a plug-and-play, recursive agent harnessing framework that addresses these challenges through two connected levels of recursion. Task-level recursion refines the current task's harness using a contrastive Propose-Probe-Compose step that reliably guides edits without ground-truth verifiers, and can be scaled further via parallel or sequential recursion. Harnessing-level recursion then extracts experience from these edits to update a continual improvement harness, enabling the agent to better adapt to future tasks. Across three professional benchmarks and four model-harness configurations, a single step consistently improves GLM-5 and Gemini 3.5 Flash over their base harnesses, yielding gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB, with further gains as the task-level recursion budget increases.

Abstract: Federated Direct Preference Optimization (FedDPO) enables privacy-preserving alignment of language models across distributed agents without centralizing sensitive preference data. However, existing FedDPO analysis assumes all clients converge to a single shared policy, which is fundamentally at odds with the heterogeneous preference distributions present in practice. We propose Personalized FedDPO (pFedDPO), in which each client maintains a small private LoRA adapter atop a shared global base model. In pFedDPO, the base model is synchronized via FedAvg each round, while LoRA adapters remain local and capture client-specific preference patterns. We prove that this decomposition replaces the gradient diversity $κ^2$ in the FedDPO bound with a residual heterogeneity $κ^2_{\mathrm{eff}}$, which is no larger than $κ^2$ when adapters absorb client-specific directions, at the cost of an additive adapter-bias term that vanishes as adapters converge. Experiments on preference datasets show that pFedDPO consistently attains a lower gradient norm than vanilla FedDPO while remaining competitive in training loss and per-client reward accuracy.

Abstract: Computer-use agents (CUAs) are rapidly improving, yet synthesizing diverse and reliable software environments at scale remains challenging. Native application environments require substantial construction and integration effort, which limits application coverage and increases the risk of application-specific overfitting. Browser-hosted environments provide limited interaction depth and workflow complexity. These limitations motivate SoftwareGym, an automated pipeline for synthesizing GUI software environments with diverse interfaces, executable functionality, and runtime-verified behavior. It generates structurally diverse designs, realizes them as runnable applications, and uses GUI evaluation with iterative repair to ensure functional reliability. Using SoftwareGym, we synthesize 182 applications across twelve categories and achieve a mean GUI task pass rate of 91.5%. We construct SoftwareGym-Bench, a 120-task evaluation suite spanning the same twelve categories, and use trajectories collected from these environments for CUA training. The resulting model improves Pass@3 from 41.39% to 56.97% on the strictly out-of-domain OSWorld benchmark and achieves 40.0% on SoftwareGym-Bench. These results demonstrate the potential of synthesized software environments as scalable and transferable supervision for CUA training.

Abstract: LLM-based agents now interact with their environments continuously, shaped by such organizational channels as user instructions, peer messages, and long-term memory. Existing safety research has examined these influences, but largely as separate agent components. How such factors jointly shape an agent's safety behavior from a unified perspective remains unmeasured. To bridge this gap, we advocate organizational behavior (OB) as a framework for studying the safety of advanced agents, reorganizing the objects of study, theoretical foundations, and experimental design around the relational structure in which agents operate. We present the first systematic formalization of counterproductive work behavior (CWB), a canonical safety-relevant subfield of OB, as Agentic Counterproductive Behavior (ACB). ACB specifies three organizational antecedents (vertical supervisor relations, horizontal peer norms, and internal cognitive structures) and maps them onto three counterproductive outcome dimensions (unauthorized disclosure, destructive operations, and production deviation). To operationalize ACB, we introduce Workerville, a controlled benchmark that manipulates organizational conditions over shared tasks, applying 16 organizational configurations to 210 tasks to yield 3,360 challenges, evaluated by human-validated agentic judges. Benchmarking 6 frontier LLMs, we find that (I) negative organizational antecedents exhibit non-monotonic amplification when combined, with the unauthorized-disclosure rate rising from 16.5% under no negative antecedent to 60.1% under two and falling back to 50.3% under three; (II) agents reproduce typical behavioral patterns predicted by human CWB research; (III) these results establish OB as a systematic framework for agent safety research, pointing toward a new research agenda.

Abstract: Federated learning (FL) typically adopts a server-client architecture, where the server aggregates clients' local models (i.e., the input) and returns the aggregated global model (i.e., the output) to clients. An untrusted server may return a tampered global model to compromise the output integrity. Some existing schemes focus on verifying the output integrity. However, these schemes mostly rely on clients pre-negotiating a set of identical auxiliary information among themselves and keeping it confidential from the server. This dependency creates a verification vulnerability: if the auxiliary information is leaked, the verification method may be circumvented. Additionally, in some privacy-sensitive scenarios (e.g., commercial federated learning), the global model may need to be kept confidential from an untrusted server. However, only a few works achieve the output privacy while addressing verification vulnerability, at the cost of prohibitive computation and communication overhead. To address these challenges simultaneously, we propose a novel provably privacy-preserving FL method called EIFL. Specifically, we adopt a two-stage aggregation and combine it with symmetric encryption to protect the output privacy. To address the verification vulnerability, we propose an efficient verification method based on vector inner product for output integrity, and a random vector generation method for clients to agree on auxiliary information. EIFL innovatively binds the auxiliary information to the output integrity, and eliminates the need to keep the auxiliary information confidential from the server. Moreover, EIFL is robust against client dropout during the verification phase through a simple resending operation. Evaluation results validate the advantages of EIFL over state-of-the-art schemes in terms of computation and communication overhead.

Abstract: Supervised fine-tuning (SFT) equips large language models (LLMs) with specialized capabilities, but often comes at the cost of forgetting the general capabilities of their parent models (i.e., the pretrained models before fine-tuning). This trade-off is especially limiting for queries that require both specialized and general capabilities. We introduce SFT-as-context, a training-free method in which the parent model uses the SFT model's response as context to answer the query. This allows the parent model to acquire fine-tuned capabilities from the SFT response through in-context learning while preserving its own general capabilities. Across 19 parent-SFT model pairs and 11 benchmarks, SFT-as-context remains close to the SFT models on fine-tuned capabilities, with gaps of only 2.2 and 2.1 percentage points on AIME 2024 and LiveCodeBench and 2.0 macro MAE on NutriBench-English, while staying within 2.2 percentage points of the parent models on general capabilities on average. Remarkably, it can solve queries requiring both fine-tuned and general capabilities, even when neither the parent nor SFT model succeeds alone. This approach also extends beyond parent-SFT pairs: responses from a small open-source SFT model can improve a strong closed-source LLM, outperforming either model alone. Furthermore, we use a Bayesian framework to derive theoretical guarantees that bound the error of SFT-as-context relative to the SFT model on fine-tuned capabilities and to the parent model on general capabilities. In addition, we visualize the attention weights and find that the parent model attends more to useful SFT responses and less to irrelevant ones, suggesting that selective attention helps the parent model use the SFT response through in-context learning.

Abstract: Tool-using LLM agents violate the policies they are deployed to enforce, often silently. Prior defenses hand-write rules, query an LLM verifier per action, or compile policies through heavyweight formal machinery. Naive compilation fails: extracted rules block the tool satisfying their own precondition, or read arguments their tool lacks. NOMOS, a four-pass compiler, turns a natural-language policy into a deterministic tool-call gate; static verification with tool-schema-level checks alone (no prover, solver, or LLM) repairs or rejects 37% (airline) and 13% (retail) of candidates, without which most shipped rules are inoperable. Replaying compiled rules over undefended transcripts flags bindings that refuse legitimate work (a development binding refused 95.9% of task-passing calls); no evaluation binding is flagged. On $τ^2$-bench the gate cuts violations of reference-encoded clauses among state-changing calls from 66.3% to 2.6% (airline) and 30.8% to 6.9% (retail), raising airline task success significantly for $2 \le k \le 4$; a 26B on-premise compilation is not significantly worse than hand-written or frontier-compiled rules. Unlike AgentDojo's shipped defenses, it reaches a zero attack success rate (ASR) on banking, where nine attack families collapse onto three structural rules. On the other three suites its ASR is at most 3.6%, from goals with no tool call to govern and one write admitted by a binding weaker than its clause; a second agent model, Llama-3.3-70B, reproduces the effect on both benchmarks. Decisions take microseconds without an LLM call, at a domain-dependent benign-utility cost; compilation runs on-premise on open-weight gemma-4-26B.

Abstract: Deployments of research agents are moving to populations of thousands that share one pool of compute, while most current systems organize one project at a time or leave the population unorganized. We argue that such a population will acquire an organization whether or not its designers provide one, so designers should provide it explicitly, and that the multi-agent systems community holds the tools to do so. We propose a society of agents, a population of persistent agents under explicit institutions, and develop it for science as a society of researchers built on six principles. Principal investigators compete for compute through requests for proposals, independent review, and grants; a human governor, the mayor, allocates resources and assigns no tasks. In a running society of ten thousand researchers, asked only to improve the pretraining of language models, one lab reported a way to reach the same quality with about 30% less compute, a result the labs that tested it do not yet agree on. We close with six open problems for the agents community.

Abstract: Many organizations adapt large pretrained models to their own tasks by fine-tuning on private data. Several of these parties often hold data for the same task and wish to fine-tune a model together without pooling that data. Federated learning (FL) enables joint fine-tuning, but reconstruction attacks on shared intermediate values (the model or its gradients) remain a privacy risk. A one-shot protocol that exchanges one encrypted contribution exposes no intermediate value. Such a protocol still gives the trained model to every participant, which is not permitted where the model is a regulated or proprietary asset. We present HE-OFT, the first cryptographically secure one-shot federated fine-tuning protocol in which no party receives the trained model. Each client fine-tunes a low-rank adapter and a classifier head on a frozen public backbone and keeps the adapter. The client uploads one encrypted head displacement, which the server combines under multiparty CKKS and never decrypts. A quorum of clients returns only the predicted label to the querier. On four text classification tasks and one vision task, HE-OFT reaches 61 to 79 per cent accuracy, against 20 to 48 per cent for a client training alone. HE-OFT keeps 85 to 96 per cent of the accuracy of a disclosed model. A test-time query takes 443.1 to 1713.1 s on one core, or 56.1 to 255.1 s with level restoration on a GPU. Restoring levels at the server cuts the traffic per query from up to 1.6 GiB to 13.5 MiB.

Abstract: Work agents operate over digital artifacts to execute professional knowledge-intensive work, requiring training environments that support long-horizon interaction and trustworthy verification. However, hand-crafted environments incur prohibitive engineering overhead that prevents environment scaling, whereas synthesis methods sacrifice workspace complexity, realism, or grounded verifiability. To bridge this gap, we introduce WorkForge, a scalable synthesis framework for constructing verifiable work-agent environments from real-world resources. Starting from expert workflows, WorkForge first identifies the resources, decisions, and deliverables required by each workflow. It then retrieves relevant real-world files and organizes them into a workspace. WorkForge inspects the workspace to extract concrete, checkable facts about its content. These factual anchors fix which task types the workspace can support and how their outcomes can be verified. Therefore, WorkForge derives each task's instructions, solution plan, and complementary programmatic and semantic verifiers directly from these factual anchors, keeping verification traceable to observable workspace evidence. Furthermore, we construct 16.7K verifiable environments across 40 professional domains, with workspaces collectively covering 60 file types. Post-training Qwen3.5-35B-A3B-Base improves GDPVal from 45.5 to 73.6 and APEX Score from 5.0 to 21.3, while enabling Qwen3.5-27B to achieve highly competitive performance and outperform strong competitors. Our analyses confirm the efficacy of the proposed method and reveal consistent scaling behaviors across both data volume and interaction horizons.

Abstract: Personal AI agents in enterprise multi-tenant deployments share a common vector store for long-term memory. Shared embedding spaces create a surface for cross-user memory leakage: a user's query can retrieve semantically adjacent memories belonging to another user through ordinary cosine-similarity retrieval, without any exploit. We formalize this as cross-user admissibility failure and evaluate it across six experiments, plus follow-up ablations, under both sparse (TF-IDF) and production-faithful (MiniLM-L6-v2) retrieval. Non-adversarial, incidental leakage reaches 70--100\% under pooled {same-team} retrieval; adversarially crafted memories achieve 90--100\% top-$k$ placement, exceeding weaker keyword-based attacker baselines, with score lifts of $+0.416$ to $+0.511$ under production-faithful dense retrieval (Config B); and end-to-end response contamination reaches 5.00/5 under a production retrieval path and 4.67/5 with Claude Sonnet~4.5, with contaminated responses often scoring as helpful or more helpful than clean ones, a gap validated against human judgment. Among three architectural mitigations, only hard post-retrieval ownership gating consistently restores the clean baseline (1.00/5) across {two generation models, at a measured latency overhead of roughly 1.4~ms per query.

Abstract: Federated Learning (FL) is a privacy-oriented learning paradigm that enables collaborative model training while keeping training data local to participating clients. However, it does not guarantee that clients submit policy-compliant contributions or that aggregators process admitted contributions correctly. Existing verifiable FL systems tailor validation rules to specific FL settings, learning workflows, and cryptographic constructions, limiting their applicability across network topologies, participant roles, and aggregation semantics. In this paper, we present PoCoFL, a policy-compliant federated learning framework that separates three aspects: (i) FL type, (ii) policy semantics, and (iii) cryptographic realisation. We provide a formalisation that captures client and aggregation requirements as policy-dependent relations. Clients prove compliance of their contributions using commitments and non-interactive zero-knowledge proofs, while aggregators prove that the recorded set of admitted contributions was processed according to the selected aggregation policy. We demonstrate PoCoFL through four formal instantiations: (i) vanilla, (ii) continual, (iii) personalised, and (iv) threshold-encrypted federated learning. We evaluate the effects of policy enforcement on the learning objectives of vanilla, personalised, and continual FL. We further implement proof-of-concept realisations of all four instantiations, demonstrating the versatility and practical feasibility of PoCoFL. Overall, these results show that PoCoFL can capture complex policy representations while remaining network-topology agnostic.

Abstract: Large language model-based agents are increasingly deployed to perform domain-specific tasks by interacting with enterprise knowledge, tools, and external services. Existing runtime guardrails primarily target prompt injection and other attack-specific behaviors under a black-box threat model, but provide limited guarantees that agents operate within their intended functionality. As a result, production agents remain vulnerable to malicious requests and out-of-domain queries that existing defenses often fail to distinguish. We present Persona Guardrail, a production-grade runtime defense framework that enforces explicit functional boundaries for customer-facing agentic AI systems through synchronous input and output validation driven by semantic allowlist and blocklist specifications. We also introduce PAGE (Persona-Aware Guardrail Evaluation), a benchmark for evaluating function-specific guardrails across benign, adversarial, and out-of-domain interactions on both user and agent turns. Compared with a generic LLM-based guardrail, Persona Guardrail improves overall accuracy from 85.7% to 95.9%, increases out-of-domain detection from 57.3% to 93.5%, and reduces the false-approved rate from 25.0% to 4.7%. Currently deployed in production, Persona Guardrail meets its latency budget while sustaining a very low false-block and false-allow rate under realistic production workloads. These results demonstrate that Persona Guardrail provides a practical, scalable, and production-ready foundation for securing agentic AI systems.

Abstract: Trajectory evaluation is essential for improving the reliability of LLM-based agents, but production use makes it expensive to run repeatedly. Modern agents generate long traces containing tool calls, observations, retries, and external outputs, while not all raw tokens are equally useful for diagnosis. We present \textit{LiteTrajEval}, a lightweight architecture for budget-bounded trajectory evaluation. LiteTrajEval derives compact domain-specific rule profiles offline, then preprocesses each trajectory online, marks heuristic failure signals, serializes it under a fixed global budget, and invokes a single rubric-guided LLM judge to produce structured diagnostic reports. Evaluated on public Magentic-One-style and $τ$-bench-style trajectory datasets, LiteTrajEval improves failure-localization alignment with human annotations by roughly 20--35 percentage points on Magentic-One and up to 23 percentage points on $τ$-retail compared with AgentRx, while reducing cost by about 6$\times$ and evaluation time by more than 8$\times$. This solution has also been deployed in our enterprise agentic platform.

Abstract: Large language models (LLMs) are increasingly deployed in enterprise, scientific, and medical applications, where agents must incorporate domain-specific knowledge and adapt from experience. Context engineering offers a practical alternative to weight updates by improving model behavior through instructions, strategies, and evidence supplied at inference time. However, adapting context online typically requires a costly trial-and-error process, while queries are often processed independently, preventing useful experience from carrying forward. Memory systems address this limitation by retaining information across interactions, but approaches that continually append information to a shared context face increasing token costs, context-window limits, and performance degradation as the context expands. We introduce a unified formulation of context optimization and show that an agent memory system update can be interpreted as an optimization update procedure over the model's context. This perspective attempts to provide a principled framework for studying memory design and its efficiency. We then propose GraphMemory, a lightweight graph-based memory that accumulates, refines, organizes, and connects reusable strategies. For each query, GraphMemory retrieves only the relevant subgraph, enabling online context adaptation without exposing the model to the entire memory. Under bounded retrieval, the amount of retrieved memory remains constant as the number of processed examples grows. Experiments show that GraphMemory achieves competitive downstream performance while using approximately 81-85% fewer memory-construction tokens than our baselines.

Abstract: Large Language Models (LLMs) have demonstrated remarkable fluency across many tasks but remain limited by their static, parameter bound knowledge and their susceptibility to hallucinating information. Retrieval Augmented Generation (RAG) addresses these issues by incorporating external retrieval into the generation process, grounding model outputs in verifiable and up to date sources. While prior surveys primarily focus on core RAG architectures and standard pipelines, recent research explores broader challenges and capabilities that extend beyond these foundational designs. This survey provides a consolidated and structured examination of contemporary RAG developments, organizing the field into a four axis taxonomy: improving retrieval efficiency, strengthening robustness and security, supporting user driven and interactive workflows, and enabling multi step or complex reasoning. We formalize key components of the RAG framework and review methods spanning dense and sparse retrieval, fusion strategies, embedding optimizations, and reinforcement learning based retrieval policies, highlighting how these advances influence practical deployment and system design. We also synthesize evaluation practices, domain specific applications, and architectural variants such as Naive, Advanced, and Modular RAG. Finally, we outline persistent challenges related to retrieval quality, reliability, domain adaptation, scalability, and explainability, and identify opportunities for building RAG systems that are more reliable, adaptable, and transparent.

Abstract: Enterprises adopting retrieval-augmented generation (RAG) face a recurring operational decision: promote, revise, or block a system version. The evidence is incomplete and the metrics come from fallible LLM judges. We report on AGO AI Quality Gate (AGO), an evidence-first quality-gate framework deployed in industrial RAG assessment engagements. AGO integrates four key components: a four-state decision model that treats missing data and judge errors as explicit outcomes; layered scoring combining deterministic checks, local guardrails, and structured LLM evaluation; a stratified beta-binomial gate that quantifies regression risk probabilistically; and a mandatory meta-evaluation protocol to validate the LLM judge before it influences decisions. Since engagement data is proprietary, we evaluate the judge layer on RAGBench, a public benchmark of 100k annotated RAG traces across 12 datasets. On identical stratified test samples (N=1200 per judge), a low-cost judge (gpt-4.1-nano) detects non-adherent answers barely above chance (AUROC 0.603 [0.570, 0.634]), despite producing flawless protocol output, while gpt-4o reaches 0.783 [0.756, 0.807] -- yet its per-domain performance still ranges from 0.62 to 0.88. A fixed-seed gate study spanning regression, no change, and improvement quantifies unsafe promotion, false-alarm cost, and improvement throughput. Under regression, the decision-grade profile reduces unsafe promotion to 22.2%-35.1%, against 29.3%-41.8% for a naive gate. These results support the design choices that judge quality must be measured per engagement and that point estimates alone are not a release decision.

Abstract: Remote multimodal models offer strong numerical reasoning capabilities over charts and speech, but sending private inputs risks exposing sensitive content. Text-only sanitization cannot directly satisfy fixed media interfaces, while identity anonymization leaves the underlying task content exposed. We introduce ReCast, an agentic plug-in framework that replaces source-specific content while preserving task-relevant relations and the required input modality. ReCast locally converts inputs into a shared textual evidence-query record, jointly rewrites entities and topics with a distilled 4B model, and substitutes values through a locally invertible, role-aware numerical map. A reconstruction agent generates and validates the required media from the protected record. The remote solver returns a program whose protected operands are restored locally before execution. On 4,000 held-out ChartQA and NMSQA examples, ReCast achieves 75.10% accuracy, retaining 92.43% of unprotected remote accuracy, while a model-based audit flags source-content leakage in 7.95% of solver-bound requests. It outperforms all evaluated local baselines, preserving the benefit of remote reasoning while reducing source-content exposure under existing media interfaces.

Abstract: Large language models (LLMs) offer strong reasoning capabilities but are often costly to access through commercial APIs, while small language models (SLMs) are easier to deploy locally yet remain weaker in reasoning. This capability-deployment gap has motivated LLM-SLM collaboration, which aims to improve SLM reasoning using LLM capabilities while preserving the deployment advantages of SLMs. Existing approaches mainly follow two paradigms. Knowledge distillation uses LLM-generated answers and reasoning trajectories to train SLMs offline, but requires parameter updates and additional training. Alternatively, online collaboration routes difficult problems to an LLM or leverages LLM-generated guidance and corrections when an SLM encounters difficulties. Although effective, online collaboration requires repeated LLM access. Moreover, the guidance produced for a particular problem is discarded after inference and cannot benefit subsequent problems involving similar reasoning states. In the paper, we focus on a more constrained setting in which the LLM is accessed only offline, the SLM parameters remain fixed, and online inference is performed solely by the SLM. To this end, we propose Reusable Latent Correction (RLC), which converts one-off natural-language guidance from a black-box LLM into persistent corrective experiences in the hidden space of an SLM. RLC stores these experiences in an external bank and retrieves them according to the SLM's current reasoning state, enabling the SLM to reuse LLM-derived corrections during inference without any online LLM calls. Experiments across multiple reasoning benchmarks and SLM scales show that RLC consistently improves SLM reasoning without parameter updates or online LLM calls. Code is available at https://github.com/ZBH031/reusable-latent-correction.

Abstract: Experienced professionals know more than just facts and conclusions. They know which cues matter, why a judgment is reasonable, and which action to take. Routine work records often leave out this tacit knowledge, making it difficult for Large Language Model (LLM) agents to use professional experience effectively. We introduce KUPAS MASTER, an experience engineering platform built around nine-layer cognitive corpus construction. It turns heterogeneous work records and practitioner interviews into traceable, reusable experience corpora for agents. Six case elements preserve the task process: context, cues, judgment, action, boundaries, and outcomes. Nine-layer cognitive corpus construction organizes tacit experience along nine extraction dimensions and stores the resulting assets in six libraries: rules, constraints, best practices, negative examples, corner cases, and skills. Semantic alignment, individual experience distillation, organizational consolidation, and cross-review preserve source evidence, conditions of use, and unresolved disagreements. The platform packages these assets into callable skills with explicit inputs, steps, dependencies, and stopping conditions, connecting experience collection to task execution and evaluation feedback. Using authorized samples from 20 randomly selected practitioners, the platform processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. The evaluation spans multiple professional domains. Under common task inputs and scoring criteria, the base model, raw corpus retrieval-augmented generation (RAG), and KUPAS MASTER agent scored 70.63, 79.75, and 89.58, respectively. The KUPAS MASTER agent improved on raw-corpus RAG in all seven scoring dimensions. The platform provides a practical path from individual tacit experience to organizational knowledge and agent capabilities.

Abstract: Federated parameter-efficient fine-tuning enables clients to adapt pre-trained models without sharing raw data or communicating the full model, but statistical heterogeneity makes a single global adapter insufficient for personalized prediction. Existing personalized methods typically use the same low-rank structure for both shared and private adaptation, overlooking their distinct requirements for aggregation and personalization. We propose FedLAFP, a role-aware framework that couples a compact, globally aggregated LoRA branch with a client-private, full-rank-capable RandLoRA branch. The shared branch provides an efficient interface for transferring common knowledge, whereas the private branch combines fixed random low-rank bases with learned scaling coefficients to provide expressive client-specific adaptation without additional communication. Client- and layer-specific mixing coefficients jointly fuse the two branches, and only the shared LoRA parameters are exchanged. A controlled linear study supports this role assignment: LoRA yields more aligned client updates and lower aggregation error, while RandLoRA more accurately recovers client-specific residuals. Experiments across four visual recognition benchmarks show that FedLAFP consistently outperforms local-only and federated LoRA baselines, achieving an average personalized accuracy of $86.93\%$ and exceeding the best baseline average by $1.30$ percentage points.

Abstract: Training models to act in accordance with an explicitly defined set of principles, or "constitution," has shown promise as a robust and transparent mechanism for AI alignment. However, the generality and flexibility of such methods remain unclear. Here, we show that constitution-consistent behavior can be distilled from synthetic corpora into lightweight objects (low-rank adapters and steering vectors). Despite never seeing a harmful request or jailbreak during training, such objects increase jailbreak defense success and measured alignment -- particularly at long context lengths and against multi-turn attacks, where they outperform both prompted and steered baselines. Subtracting control-trained from constitution-trained objects further accentuates these effects, yielding defenses we call "constitutional adapters" (CAs). CAs can be trained on a base model, transferred zero-shot to its post-trained checkpoint, and scaled at inference time to predictably trade off defense for benign compliance. Taken together, these results recommend CAs as a lightweight, portable, and tunable lever for mitigating misalignment and misuse in API deployments.

Abstract: As LLMs increasingly act as delegated agents, they are expected to protect principals' interests when interacting with external parties. Standard alignment objectives, such as helpfulness, harmlessness, and honesty, do not specify how agents should protect principals' strategic interests under delegation. We formalize Agent Loyalty as an information-control property requiring agents to prevent Exploitable Information Leakage (EIL) and resist Manipulative Information Uptake (MIU). We introduce LoyalAgent-Bench, comprising 10,298 samples across 42 subscenarios and six domains, and an online GRPO framework that trains against a LLM opponent to generate mechanism-specific reward signals. Experiments show that loyalty is not guaranteed by general capability or existing alignment, with measurable EIL and MIU gaps under zero-shot evaluation, while our trained 8B models reduce per-response leakage in single-turn exchanges by 31-44pp and improve task utility, evidence faithfulness, and decision accuracy by up to 11pp, 49pp, and 29pp, respectively. For the trained Qwen3-4B model, no degradation is observed on out-of-distribution benchmarks in math and narrative reasoning.

Abstract: While Large Language Models (LLMs) advertise million-token context windows, reasoning quality often collapses as inputs grow -- a phenomenon termed context rot. This failure stems from a structural entanglement in monolithic architectures, where the massive search burden of contextual grounding exhausts the representational capacity needed for complex reasoning. To resolve this, we propose Grounding-Reasoning Disaggregation via DIStributed long COntext scaling (DISCO). Inspired by distributed computing frameworks like Apache Spark, DISCO partitions long context across a fleet of Worker LLMs dedicated exclusively to parallel, localized grounding. A central Driver LLM, trained via Reinforcement Learning (GRPO) to optimize planning, orchestrates execution by dynamically mapping queries into atomic extraction tasks and reducing the gathered evidence to synthesize a final answer. By isolating reasoning from raw context noise, DISCO effectively eliminates context rot. On RULER-QA (1M tokens), it maintains 78.4% accuracy where standard baselines collapse. Furthermore, it outperforms full-context models by up to 9.8 points on LongBench v2 and matches frontier models like Gemini-3-Pro-Preview while reducing inference costs by over 80%, establishing a highly efficient paradigm for robust long-context inference.

Abstract: Retrieval-augmented generation (RAG) lets language models answer questions more accurately by consulting relevant documents. Many valuable collections, such as medical records, cannot be pooled because of privacy rules. Federated RAG leaves each collection with its owner, or node, which scores candidate answers from its own documents; a central hub combines the scores. Some nodes, called Byzantine, may be compromised, faulty, or misled by instructions hidden in documents, and report arbitrary scores. Conformal prediction returns a set containing the correct answer with a chosen probability, using a cutoff set in a calibration step on questions with known answers. An unknown group of nodes, no larger than a declared bound, may misreport both in this step and at query time. Existing methods assume every node is honest or protect only the calibration step. We observe that the honest nodes are the same in both steps. The hub therefore has all nodes score the same calibration questions, and keeps a candidate only if some plausible group of honest nodes, using its own scores in both steps, would keep it. We prove that the resulting sets contain the correct answer with the chosen probability in finite samples, whatever the Byzantine nodes report. No method using the same information can return smaller sets without risking the loss of an answer the honest nodes support. If nodes fail at random, the guarantee weakens only by the probability that more nodes fail than declared. In simulations, on real question-answering tasks including medical exams, and with language models as nodes, some hijacked, our sets reached the target whenever no more nodes misbehaved than declared, while plain averaging could miss it. They were also clearly smaller than those of simpler methods with the same protection, most of all when the declared bound was generous, so a cautious bound costs little.

Abstract: Comparative Process Reward Models (PRMs) provide critical step-level guidance for autonomous web agents by evaluating state-conditioned preferences between candidate actions. However, existing preference training data synthesized via multi-policy sampling suffers from a severe scarcity of Grounded Minimal Contrastive Pairs (GMCPs)-where competing candidates target genuine on-page elements with identical action types. In representative baselines preference data (namely, WebArbiter), GMCPs account for merely 24.19%, biasing PRMs during training to rely on shallow shortcuts (such as element hallucinations and action type mismatches) rather than acquiring genuine contextual decision semantics. To address these challenges, we propose SURFPRM, a graph-guided process preference synthesis framework for comparative Web PRMs. SURFPRM structures web demonstrations into a persistent Interaction Element Graph that acts as an environment-grounded negative action proposal mechanism, systematically synthesizing contrastive negative actions across spatial, temporal, and spatiotemporal confusion axes. This elevates the GMCP proportion from 24.19% to 74.60%, producing the curated SURFPRM-DATA dataset. Across six open-source backbones (3B to 9B parameters), PRMs trained on SURFPRM-DATA outperform baseline-trained models on average on WEBPRMBENCH and rival leading proprietary LLMs. In downstream reward-guided trajectory search on WEBARENA-LITE, SURFPRM provides step-level guidance for both GPT-4o (+14.21%) and GPT-4o-mini (+12.83%) policies, yielding substantial improvements in complex web task success rates.

Abstract: Language models are often evaluated on curated benchmarks that underrepresent the complexity of enterprise deployments. We introduce HARDEN, a constrained evolutionary search method to adapt the input of existing evaluation cases into more challenging variants while keeping their expected outputs fixed. HARDEN searches along generated domain-specific complexity axes while enforcing feasibility constraints such as preserving task semantics, realism, and execution validity. Across FinQA, PubMedQA, and ContractNLI and three Qwen3.5 model scales (35B-A3B, 122B-A10B, and 397B-A17B), HARDEN reduces task-model accuracy by 22.7% on average and by up to 49.9% relative to single-pass baselines using the same feasibility checks. These results show that evolutionary search can produce substantially harder valid evaluation cases.

Abstract: While graph-based and hypergraph-based Retrieval-Augmented Generation (RAG) significantly mitigate hallucinations in Large Language Models (LLMs), existing structure-based RAG systems typically adopt static traversal strategies regardless of the query complexity. We identify this ``static retrieval fallacy'' as a primary source of computational redundancy for simple queries and cognitive context gaps for complex reasoning tasks. To balance reasoning quality and inference efficiency, we propose Asymmetric Dynamic Routing (ADR), an intent-conditioned retrieval framework operating over hierarchical knowledge graphs. ADR employs a lightweight structured classifier to dynamically dispatch queries among three asymmetric topological traversal operators: localized fact anchoring, bottom-up adjacency diffusion, and top-down insight grounding, which collectively enable bidirectional information flow across hierarchical knowledge layers. Extensive empirical evaluations across five domain-specific corpora demonstrate that ADR maintains strong reasoning performance while reducing prompt token consumption by up to 48.7\% and end-to-end query latency by 45.3\%, yielding a favorable quality--efficiency trade-off for query-adaptive Hypergraph RAG.

Abstract: Harnesses, the products that run AI coding agents, are multiplying, and enterprises are rolling them out to their employees: what started as pilots with a few hundred seats is scaling to tens of thousands. Most enterprises do not build these harnesses but buy them from large vendors, such as Anthropic's Claude Code or OpenAI's Codex. A harness decides which model answers, what the model reads, how the prompt cache is used and which subagents run, so it picks the rate on the price sheet and sets the volume bought at it. Enterprises that keep a proprietary or untuned harness at its defaults inherit these choices and their bill. We build a fast, customisable router in which Jev, a classifier with calibrated probabilities, labels every prompt against a bring-your-own taxonomy of agentic requests. Because one user turn is many requests over a prompt cache that belongs to one model, the router moves work only where no running conversation has to rebuild its cache: at session start, in side lanes and at subagent launch. From the price sheet we derive when a mid-task switch pays back, and a crossover: on long tool-heavy sessions the highest-priced model costs less than the next tier, as repricing about 10,000 real sessions from public datasets confirms. In an emulated enterprise of 10,000 seats with user behaviour taken from these datasets, the router recovers 14 to 21% of model spend at Anthropic's list prices of 21 September 2026, \$3.3M to \$5.0M a year. The paper also maps the risks across twenty harnesses, prices the dependence on one vendor's models, and proposes a control plane that enterprises can run from within, starting now, with a ladder for deciding later whether to own the harness.

Abstract: Contract inference requires multiple judgments about a shared document, but aggregate accuracy can conceal changes in the individual decisions. Repeated agreement is also insufficient: a model may consistently return the wrong answer. In this paper, we compare Jev with nine language models on ContractNLI, evaluating inference cost, response time, average correctness, and correctness across repeated request conditions. Controlled comparisons vary hypothesis visibility, requested outputs, and output order while keeping the contract and target judgment fixed. Jev has the lowest cost and median response time among the evaluated configurations, while hosted language models achieve higher baseline accuracy. Rankings by baseline accuracy differ from rankings by correctness across every condition and repeat, although small differences in the latter do not establish a general stability advantage. Development diagnostics further reveal compensating corrections and regressions, as well as persistent errors. These findings motivate evaluating cost and response time alongside whether individual judgments remain correct as the request configuration changes. Code: https://github.com/ZF-Utokyo/Jev-Benchmark

Abstract: On repeated work, agents are inconsistent. We ran 42 tasks three times each and found that, depending on the model, 38% to 74% returned answers that did not agree. Consistency is what a buyer, an auditor, or a regulator requires, and agents do not have it. They are wasteful too: 95.3% to 97.2% of what an agent generates goes to re-deriving a plan the system already knows. We propose skill habit formation. An agent mines its own execution history for candidate skills, deterministic variants that compete against the incumbent rather than replacing it. A candidate declares the region of input space it claims, so the common case runs as a script and the rest falls through to reasoning. Four gates of ascending cost admit candidates; the central one tests a candidate's execution trace against a retained reference, within a tolerance measured from that reference's own run-to-run variability. On text-to-SQL, three of four reasoning arms reproduced their own output on 11 to 13 of 42 repeated questions and the fourth on 26 of 42, while a habit-formed variant reproduced on all 456 dispatches we repeated and was non-inferior to every arm it replaced (p<0.0001). It also used 14% to 56% fewer tokens, turning net positive after 7 to 53 reuses. We measured what this costs in accuracy. The guard admitted work it should have deferred on 2.6% of natural paraphrases and 26% of inputs near its boundary, and 11 of 13 such failures were invisible to the trace-conformance gate at any threshold. Deterministic errors repeat exactly: a bad habit is as reliable as a good one, and that is the price of the property that makes the system auditable. Separating routing from parameter extraction raised end-to-end accuracy from 0.888 to 0.952 at 43% of the cost.

Abstract: Tool-calling LLM agents are increasingly deployed in enterprise applications. However, effective evaluation and optimization require high-quality, diverse task datasets that are often difficult to obtain due to privacy and other constraints. Existing synthetic task generation methods often produce generic tasks that ignore an agent's underlying state or database and fail to reflect real-world usage diversity. We propose EdgeGen, a synthetic task generation framework that extracts compliance rules from an agent's specification and uses them to generate database-grounded edge-case tasks designed to violate these rules. When combined with existing synthetic data generation techniques, EdgeGen enables agent improvement through finetuning and harness optimization. The resulting pipeline forms a fully automated closed-loop system that requires no human annotation. Finetuning on data generated by EdgeGen yields a consistent mean progress improvement of 2 percent to 42 percent on tau2bench airline domain, while other baseline methods show degradation for some models. On the other hand, for harness optimization, our method shows a mean progress improvement of 10 percent and 30 percent over the human-curated and base harnesses, respectively, for the Gemma-4-e4b model.

Abstract: Two transformative classes of AI systems are reshaping how organizations operate: \textit{digital AI}, which reasons over enterprise knowledge to power chatbots and agent workflows; and \textit{physical AI}, which learns to control robots and autonomous systems from video, gameplay, and sensor telemetry. Both face the same foundational bottleneck: raw knowledge at scale, spanning heterogeneous modalities, locked in private corpora that existing AI infrastructure cannot access reliably or efficiently. We propose \textit{Universal Knowledge Perception (UniK)} as a common platform for both classes, covering the full knowledge lifecycle (ingestion, enrichment, indexing, retrieval, and continuous evaluation) across modalities from rich text and video to molecular data and sensor telemetry. We present UniK, built on Polymath Retrieval (multi-index fusion over automatically enriched indices) with no task-specific fine-tuning. Across five digital AI domains (medical literature, open-domain QA, chemistry, legal video proceedings, and government open data) UniK combined with an open-source 70-billion-parameter model consistently matches or outperforms frontier proprietary LLMs that are orders of magnitude larger: 76\% RAG accuracy on government data versus 47\% for GPT-5; 77.9\% on medical QA without fine-tuning; topping all open-source chemistry pipelines. We show that the same infrastructure directly addresses the data curation, indexing, and retrieval challenges facing physical AI world model training, where the knowledge problem is harder but structurally identical.

Abstract: Agentic AI systems built on large language models can plan over multiple steps, use external tools, retain information in memory, and coordinate with other agents. These capabilities make them more useful than static language models, but they also introduce new security and operational risks. Untrusted content from websites, emails, documents, and databases can enter the same context as system instructions; persistent memory can carry compromised information across sessions; and access to external tools can turn an incorrect model response into a consequential real-world action. This article reviews the trustworthiness of agentic AI across five interconnected dimensions: safety and robustness, alignment and human oversight, transparency and auditability, privacy and data governance, and regulatory compliance. It organizes key failure modes, including indirect prompt injection, backdoor triggers, goal misgeneralization, memory contamination, and cross-session data leakage, into a unified taxonomy. It also examines major mitigation approaches, such as instruction hierarchies, context isolation, spotlighting, process-based supervision, constrained tool use, and privacy-preserving memory, while distinguishing techniques supported by empirical evidence from those that remain largely conceptual. Building on this analysis, we introduce the Trustworthy Agent Development Lifecycle (TADL), a six-phase framework covering specification, design, training, evaluation, deployment, and monitoring. For each phase, TADL identifies relevant trust activities, expected evidence, and risk-based decision gates. Although TADL has not yet been empirically validated, it provides a structured foundation for developing and evaluating more secure and accountable agentic systems. The article concludes by identifying gaps in current benchmarks and outlining priorities for future research.

Abstract: Frontier language models now produce professional deliverables that expert graders judge to match human work on a substantial share of economically valuable tasks, yet most enterprise GenAI initiatives fail to show a measurable business effect and a large fraction of agentic projects are expected to be cancelled. We argue that this is substantially a measurement problem: public benchmarks answer "what can the model do?", whereas a deployment decision requires "is this workflow fit, reliable, safe and worth scaling - here, on our data, under our controls?". We present EnterpriseVal, a use-case-level evaluation system that closes this gap. It comprises (i) a formal specification of the use case and of the frozen socio-technical configuration under test, model, prompts, retrieval, tools, guardrails and human oversight, with an autonomy level and consequence tier that jointly set the required evaluation intensity; (ii) a metric catalogue spanning fidelity, utility, efficiency, reliability, assurance and oversight; (iii) a grading protocol that scales blinded expert judgement with calibrated LLM-as-judge scoring through prediction-powered inference; (iv) a two-tier threshold gate, stated as an executable algorithm, that maps metric vectors with confidence bounds to REJECT/CONDITIONAL/SCALE decisions; and (v) a value-and-risk model in which the reviewer catch rate is a measured parameter. We report a pilot across three workflows in a global bank. In credit-memo drafting, human-graded citation precision reached 88% and hallucination rate 1.6% for the best model against gates of 70% and 5%; in procedure transformation, analyst refinement effort fell from an estimated 27.4 to 2.9 hours per document. We separate established results, documented pilot evidence, the proposed system and open hypotheses, and specify the experiments required for full validation

Abstract: With the widespread industrial deployment of Large Language Models (LLMs), DeepSearch has emerged as the dominant paradigm for resolving complex user queries. It typically operates through an iterative closed-loop workflow consisting of planning and reflection, information retrieval, and answer generation. However, existing reward models (RMs) and evaluation benchmarks are primarily designed for static single-turn tasks, failing to capture the full-pipeline complexity of DeepSearch workflows. To address this limitation, we propose F2DR, a fine-grained full-pipeline DeepSearch reward framework. F2DR evaluates DeepSearch workflows across three dimensions: Content, Trajectory, and Answer, enabling comprehensive process-level assessment. We further construct DeepSearch RM-Bench, a dedicated benchmark for evaluating RMs in DeepSearch scenarios. Extensive experiments demonstrate that F2DR achieves significantly higher evaluation consistency than self-evaluation-based baselines, while DeepSearch RM-Bench exhibits strong discriminative capability across existing open-source RMs. We will publicly release the complete DeepSearch RM-Bench dataset soon.

Abstract: As corporate AI adoption continues to grow, enterprise-grade LLM agents are being deployed into sensitive contexts such as hiring, healthcare, and finance. In these contexts, compliance with rules specified in an agent's system context is a first-order legal concern. Currently, no evaluation framework systematically measures which LLM models tend to violate compliance rules, especially under pressure from a persistent user, a hurried manager, or circumstances where violation is convenient or attractive. We introduce PACT (Pressure-Applied Compliance Testing), a benchmark for rule-following under pressure in AI agents assisting employees in daily tasks across twelve regulated enterprise domains and forty-eight scenarios, each set in a realistic multi-turn conversation. Each benchmark item pairs a standing rule against a rule-violating shortcut, and applies a battery of pressures across different wordings and system-prompt modes. We construct PACT component by component under strict LLM-as-judge auditing to ensure samples are unambiguous, ungameable, and realistic enough to avoid eliciting evaluation-aware behavior. We use PACT to profile LLM compliance across six complementary metrics that create a holistic picture of an AI assistant's robustness under pressure and throughout multi-turn conversations, its transparency, and ability to correctly discern where a rule applies. We aggregate this profile into PACTScore, a reliability-weighted compliance rate over all items and modes. Our results across 22 common LLM models spanning multiple providers and sizes show substantial variability in compliance across models and metric dimensions. Even the strongest assistants mis-apply a rule on 6 to 10% of items, and ordinary user pressure raises the violation rate by 65% on average. PACT highlights compliance risks in LLM assistants, motivating guardrails and careful model selection.

Abstract: Real-world agents fundamentally require persistent non-parametric knowledge for dynamic reasoning, i.e., long-term memory and retrieval-augmented generation. While graphs have shown reliable advantages in providing structured evidence, the sparse graph representations naturally restrict machine readability and semantic density required for complex agentic workflows. Driven by this limitation, the entire industry is witnessing a paradigm shift from traditional sparse graphs to LLM Wiki, an agent-native knowledge representation that couples dense document contexts with markdown files containing multi-layered topological linkages. However, parameterizing such rich semantics is challenging to encode dense textual contexts using traditional sparse graph embeddings. Moreover, learning LLM Wiki with existing graph encoders could overwhelm distributed system overheads that hinder deployment in large-scale commercial scenarios. To this end, we propose a novel paradigm Wiki Foundation Model, i.e., WFM, tailored for scalable, agent-native representation and retrieval. Specifically, (i) we formalize a Wiki Graph schema that seamlessly bridges fine-grained structures with dense contexts, maintaining explicit topologies alongside continuous semantics; (ii) A query-conditioned attentive aggregation is tailored for rich wiki message passing and explicit attention variance regularization; (iii) We engineer an infrastructural NCCL boundary exchange protocol that hoists static partition indices and leverages fixed-shape GPU-to-GPU collectives, bypassing CPU serialization and memory copy overheads. Extensive evaluations across five long-term agent memory and multi-hop reasoning benchmarks demonstrate the remarkable performance of WFM, while achieving a 10.5 times training acceleration on distributed clusters.

Abstract: We present ORQA, a method for testing occupation-level knowledge in large language models. Prior methods either map abstract LLM skills to occupations via task definitions or utilize expert knowledge which is difficult to obtain at scale and expensive. ORQA complements both of these methods by connecting O*NET occupations to trusted occupation-specific websites (such as regulatory agencies, licensing bodies, professional organizations, and government publications) and converting these into source-traceable question-answer pairs. A combination of an automated pipeline and human review produces a set of high quality questions about occupations. The question set created via our method covers 116 occupations from all 21 major groups in the SOC, with 480 questions sourced from 187 different websites. Each question is designed to probe a real-world skill question that is relevant to the occupation in question. We test 15 state-of-the-art frontier and open-weight models via this method. Claude Opus 4.6, GPT-5.4 and Claude Sonnet 4.6 all perform the best at approximately 58-62% while smaller open-weight models achieve approximately 33-41% performance. Performance varies significantly across occupations. Healthcare-related occupations achieve the highest performance (78%) while Office and Administrative Support achieve approximately 40%. Performance on individual occupations (e.g. Sheet Metal Workers and Fish and Game Wardens) is essentially zero. We also find that open-ended questions and weighting by wage bill do not significantly affect the ranking of models on this benchmark. We believe that leveraging existing trusted occupation-specific information to test LLM knowledge in professional domains may be a scalable and useful method for evaluating occupation-level AI performance in the future. Results and data are available at orqabench.org.

Abstract: Enterprises deploy systems, not checkpoints. Usable capability depends jointly on weights, serving route, precision, output contract, and harness, yet all 18 audited benchmarks score advertised model identifiers. We treat this as measurement error and give a protocol that makes it reportable. It has three parts. A gold-blind capability-binding preflight verifies that a route can execute the evaluation contract before any task reaches it; a reliability-inclusive first-pass scoring rule keeps failure in the score while keeping unsupported capability out; and adjudication is structurally score-blind. We call the protocol IB2 and release its algorithms, classification tables, request contract, and manifest schemas. Its reference instantiation, 128 locked tasks and 987 assertions over document, spreadsheet, chart, tool and database work, stays sealed: the procedure is the artifact, not the corpus. Across eleven systems, four results. Capability availability is measurable: two complete single-route runs on identical weights later failed distinct predicates of the finalized binding gate, while a third passed that gate before a fresh run. The advertised identifier exposed neither limit. Discrimination is not uniform: four of seven suites saturate under a six-system band, with the spread almost entirely from governed database work and multi-tab joins, so we report interval-backed resolution groups, not ranks; two of the nominal five-label output's four cuts fail multiplicity adjustment. Serving-arm choice moved one declared revision and precision from 77.38 to 82.54, paired interval [0.11,10.60], though the arms differ in access mode, harness generation, and the serving tool-call parser, and harness generation is a property of our evaluator, not any endpoint. Excluding failed responses from denominators changes the point ordering, so reliability inclusion changes a conclusion, not its wording.

Abstract: Open-source large language models (LLMs) are increasingly competitive with closed-source models while offering transparency and the ability to run inference without exposing user inputs to a service provider. However, running large-scale models locally requires substantial computational resources. In practice, users may still resort to a third-party provider, giving rise to privacy and correctness concerns. Existing solutions that address these problems often impose substantial server overhead or introduce additional trust assumptions. In this paper, we present Maverick, a novel approach to private and verifiable LLM inference based on a protocol for delegating matrix-vector multiplication, a dominant operation in LLMs. At its core, Maverick provides, to our knowledge, the first information-theoretically sound verification protocol for matrix-vector multiplication delegation with transparent preprocessing, efficient (batch) verification, and virtually no server overhead. We combine this verification primitive with LPN-based pseudorandom masking to provide input privacy. We implement our matrix-vector delegation primitive and use it to build an end-to-end prototype of Maverick, which we evaluate on Qwen3-4B by measuring throughput in tokens per second. We evaluate client configurations with 1-8 threads. With one client thread and a CPU server using up to 128 threads, Maverick achieves throughput gains over local inference of up to 17x when privacy masks are generated online, 45x when they are precomputed, and 44x when only verification is required. With four client threads, the corresponding gains are 13x, 18x, and 17x. When server computation is no longer the bottleneck, client-side microbenchmarks with simulated network delay show speedups of 12x-20x, 34x-135x, and 38x-157x.

Abstract: AI systems become more useful when they can adapt to the people using them, but in multi-agent systems, useful context learned by one agent often remains unavailable to others. We present kernel-managed shared memory, a system-level abstraction in which specialized agents write structured, tagged memories while the agent-system kernel, not individual agents, governs retrieval, privacy enforcement, and prompt injection. We implement and evaluate this design on AIOS and compare it against three alternatives across three assistant models (GPT-4o, Llama-3.1:8B, Qwen-2.5:7B) and 1,800 total trials. Against an unmanaged external memory backend (Mem0) using identical underlying storage, kernel-managed retrieval and injection improve personalization scores by 2.4-4.0 points on a 5-point scale (e.g., 1.05 to 4.69 profile usage on GPT-4o), with every comparison significant at p < 10^-18. Against standard retrieval-augmented injection, gains are similarly large and consistent across all three models. Against full, unfiltered context concatenation, a soft ceiling on available context rather than on response quality, kernel-managed injection statistically matches performance on two of three models and shows a small, model-specific deficit on the third, while using substantially shorter prompts: end-to-end latency is 15-61% lower across all three models, with corresponding reductions in per-call token usage and inference cost. These results indicate that centralizing memory management in the agent-system kernel, rather than leaving retrieval and privacy enforcement to individual agents, delivers most of the personalization benefit of unconstrained context at a fraction of its cost.

Abstract: Fine-tuning large language models (LLMs) on domain-specific data is essential for downstream adaptation. In many deployments, a participant cannot hold the complete model locally. This happens because the model owner keeps the full model proprietary, or because the participant lacks sufficient compute resources. Split Learning (SL) addresses this by partitioning the model between the participant and a server so that only a small portion runs locally. When the underlying data is additionally distributed across multiple institutions with privacy requirements, Federated Learning (FL) further enables collaborative training across participants by sharing only model updates instead of raw data. In this combined setting, each client transmits intermediate activations to the server, and for LLM fine-tuning, this exchange poses an inherent privacy paradox. The autoregressive nature of LLMs causes the transmitted activations to leak the input, and existing perturbation-based defenses are fundamentally ineffective in this setting. We address this leakage through a learned obfuscate-and-recover scheme that protects participants' private datasets while still allowing an independently deployable model to be trained on the server side. Experiments demonstrate that our approach achieves strong privacy protection with modest utility loss and system overhead, making split-based federated LLM fine-tuning practically viable.

Abstract: Long-running AI agents may read state, reason, wait for tools or human approval, and perform an external action much later. The state that justified the action can change in the meantime. For example, after an agent proposes an 80 GBP refund under a limit of 100, a customer-name change affects only presentation metadata, a new limit of 90 still permits the refund, a limit of 50 invalidates it, and a refund issued by another worker must prevent a duplicate. Standard optimistic concurrency control and version checks can detect that previously read state has changed, but by themselves do not determine whether that change invalidates the pending action's justification. We call any detected version change a version conflict; when that change invalidates the action's justification, it is also a decision conflict. ATR records the explicit, executable conditions that justify a pending action and rechecks only the conditions affected by a change before releasing the external operation. It can retain the action, refresh non-decisive metadata, require replanning, or block execution; a target-side transaction or compare-and-set binds checked state to commit. Across 210,000 controlled executions over 15 mutation cases, ATR matched every developer-specified outcome with no false allows or blocks. In ten durable SQLite checkpoint/resume cells, it evaluated 0.6 conditions per change versus 6.0 for FullScan. At 4,093 recorded reads, ATR took 9.3 microseconds versus 2595.9 microseconds for FullScan. These deterministic results establish controlled feasibility, not production generality or automatic extraction of the required conditions.

Abstract: Multi-agent LLM pipelines increasingly span organisational boundaries; when a fault surfaces, someone must determine where it entered. The artifact available is rarely a full execution trace: it is the reports each agent filed, and a filed report can state a conclusion alongside its observations. Using a pre-registered, institutionally partitioned pipeline of six agents with process-level information boundaries, balanced defect injection and matched clean twins (345,600 requests per chain model, two models), we first report that our pre-registered hypothesis -- that collective responsibility framing degrades escalation with chain length -- is not supported. The layer nevertheless fails asymmetrically. It originates almost nothing: zero allegations across 7,996 clean episodes where every agent stayed silent. It filters upstream error poorly, naming an innocent party in 34.4% and 62.6% of clean episodes where an agent raised a false alarm. Conditional on no agent proposing the true origin (59.5% of episodes on one chain model), an auditor reading the reports recovers it in 4.1% of cases -- below a uniform guess (20%) and the best fixed-link accuser (31.0%) -- while reaching 60.3% from the raw documentation of the same episodes. Deleting one clause, the field carrying the agents' own conclusion, isolates the cause at constant observations: accuracy rises to 45.2% (+41.2 pp, 95% CI +35.3 to +46.9) and adherence collapses from 94.4% to 3.4%; where the suggestion was correct the same deletion instead costs accuracy, 70.5% to 55.7%. The harm replicates on two frontier auditors in four conditions out of four (+8.5 to +39.0 pp) and in a second domain (+47.7 and +61.1 pp), where the cost disappears. The net effect is governed by upstream reliability together with both conditional magnitudes. An accountability layer needs evidence sufficiently independent of the conclusions it verifies.