enterprise ai developers

Abstract: Agentic AI drives significant productivity gains across diverse domains, but concurrent long-context prefills from multiple agents can saturate GPU compute resources. Although CPUs offer a viable alternative, existing CPU attention kernels fall short in adapting to multi-tier cache hierarchies and lack effective thread scheduling. To address this, we introduce CACPU, a high-performance attention kernel featuring tiered cache blocking and an in-kernel task dispatching mechanism designed for causal full-sequence prefilling. CACPU achieves up to a 2.01x speedup over PyTorch SDPA, FlexAttention, and SGLang attention kernels. Integrated into SGLang, it reduces Time to First Token (TTFT) of Llama-3.1-8B by up to 1.74x while preserving model accuracy.

Abstract: Recent advances in large language models (LLMs) have spurred growing interest in temporal knowledge graph question answering (TKGQA). Most existing methods rely on semantic retrieval augmented generation (RAG), which suffers from fundamental limitations in complex temporal reasoning. On the one hand, semantic-based RAG methods struggle to capture the multi-hop causal knowledge required for complex temporal reasoning. On the other hand, these methods introduce excessive redundant knowledge, trapping LLMs in spurious reasoning chains that potentially result in erroneous answers. To address these problems, we propose a Temporal Causal Chain-guided RAG (TCC-RAG) to retrieve a concise yet sufficient set of causal knowledge to support complex temporal reasoning. Specifically, we reformulate the temporal knowledge graph into a multi-view causal cue graph, with causal cues inducing connections that explicitly encode temporal evolution and causal dynamics. Based on the causal cue graph, we construct a query-specific temporal causal chain to guide the retrieval process, enabling the model to focus on a concise set of causally relevant knowledge. Finally, we introduce a causal evidence filter to refine the retrieved knowledge, ensuring its causal consistency with the underlying temporal logic of the query. Experiments show that TCC-RAG achieves substantial performance gains of 10.72% and 8.40% on two TKGQA benchmarks, and delivers a remarkable 62.97% improvement for multi-step temporal reasoning.

Abstract: The Persona Selection Model (PSM) posits that post-training selects a persona the model learned during pre-training. Work on subliminal learning has shown that models can impart covert signals (like "love of owls") into semantically unrelated outputs like strings of numbers, and that students trained on these outputs inherit those traits. In this work I show that by evaluating the likelihood of such outputs under the base pre-trained model with an appropriate header, a method I call persona-header likelihood evaluation (PHLE), one can decode these subliminal signals and classify between them. For example, PHLE distinguishes number sequences generated by a post-trained model prompted with "You love owls" from those prompted with "You love eagles", and distinguishes answers from a model with a secret desire to harm the user, filtered to remove any noticeable misalignment, from those of a model with no system prompt. In both cases prompted classifiers, including the generating model itself, remain near chance. The signal partially persists across model families, with the Qwen base model reading signals generated by Olmo and vice versa. This is evidence for a non-trivial prediction of the PSM, and the method may be practically useful for auditing the hidden objectives of models.

Abstract: Conventional federated learners implicitly optimize for global means, creating a "one-size-fits-all" paradigm that inevitably suppresses the minority under cross-site heterogeneity. To bridge this critical gap, we present Fed-Universe, a generalizable Semantic-Geometric-Topological-Human (S-G-T-H) architecture that transforms passive averaging into an active negotiation for decentralized alignment. The S-layer deploys an edge-capable LLM as a semantic surrogate to translate heterogeneous inputs into standardized patient profile prompts. The G-layer enforces geometric quality assurance (GQA) via a mapping f: Theta -> Q from the statistical parameter space to a physically interpretable quality domain, employing a cosine similarity gate to retain minority updates based on physical geometry rather than statistical proximity. The T-layer executes topological Pareto control, identifying a sensitivity-based knee point (lambda_k) to balance multi-objective aggregation without majority degradation. Finally, the H-layer operationalizes a cognitive twin dynamic to mirror the user's real-time psychological state, dynamically adapting human-computer interaction modes via active intent alignment. To ground this theoretical architecture empirically, we validated the S-G-T core on a binational synthetic clinical network simulation. This proof-of-concept demonstrated a "zero-sum escape": a minority node (<3% data volume) achieved utility parity (loss reduction from 0.857 to 0.340) alongside dominant hubs. Our Fed-Universe framework is being expanded along two focused pillars: institutional multimodal alignment (Fed-Ultra) and individual cognitive intent alignment (Fed-Human).

Abstract: Large language model agents that use tools could take on part of the electronic health record and desk work that occupies physicians' time, but in many hospitals the medical information systems run inside closed networks, so the model and the agent harness must also sit inside. We designed, implemented, and deployed LLENS, a clinical support AI agent platform that runs inside the closed network of Hokkaido University Hospital on one host with 8 NVIDIA H200 GPUs, the open-weight model Kimi K2.6 (~1 trillion parameters), and a web application rebuilt from Open WebUI. The harness has 10 components, including page image reading, full-text attachment input, a path-addressed document tree without a vector index, a network-isolated sandbox, artifact return, self-checking of generated output, subagents, and saved procedures. Kimi K2.6 was chosen from 3 open-weight candidates and scored 480/500 on the 2026 Japanese Medical Licensing Examination; it generated 119 tokens/s for one user and served 16 concurrent requests at 908 tokens/s in total. On 180 questions over hospital documents (3 seeds each), mean accuracy was 10.7% with no documents, 81.3% with fixed retrieval (top 10 chunks of one lexical search), and 97.2% with interactive access, in which the model reads with document tools and the sandbox (15.9 percentage points above fixed retrieval, 95% CI 10.9-21.3); the difference lay in filtering/counting questions (16.7% vs 87.8%). Five scenarios with synthetic data ran on combinations of the components: needlestick injury response, batch adverse event grading, scanned referral letter transcription, falls data aggregation and plotting, and investigation of 3 topics by subagents. The platform has run on the hospital network since September 14, 2026.

Abstract: Reinforcement learning (RL) is an effective way to strengthen the long-context reasoning of large language models (LLMs), but it needs training data that are long, diverse, challenging, and verifiable. Human annotation of such data is expensive, naturally occurring data are too scarce to sustain RL, and existing synthesis methods depend on external corpora or stronger teacher models, which become harder to obtain as the trained models get stronger. Recently, RL from zero data, in which a model learns from tasks it proposes itself, has achieved excellent performance, but existing methods target short-context tasks. Therefore, we propose $\textbf{LongZero}$, a method that lets an LLM improve its own long-context reasoning without any external data. Instead of writing questions and answers over existing documents, LongZero first has the model synthesize the question and answer with its supporting evidence in a compact structured form, and only then separately expands the evidence into a long multi-document context. LongZero combines these self-written multi-hop reasoning tasks with self-written multi-round retrieval tasks and trains the model with GRPO. We evaluate LongZero on two LLMs across five long-context reasoning benchmarks and compare it with RL on data synthesized from external corpora. Experimental results show that LongZero can generalize to real tasks, and outperforms both baselines, suggesting a viable and promising path to improving long-context reasoning of LLMs.

Abstract: Chain-of-thought (CoT) monitoring is one of the few practical tools available for overseeing capable AI systems, but it is fragile: a model's reasoning is only useful to a monitor if it remains coupled to what the model actually does. This coupling can be degraded during training through, for example, CoT compression, optimization against a monitor, and out-of-context generalization from synthetic documents. In this work, we propose an unsupervised post-training objective based on maximizing mutual information between CoT and the output for improving CoT monitorability. Across three model organisms with degraded CoT, it recovers most or all of the lost monitorability. Applied to undegraded policies, it further improves judged monitorability, improves prediction of unsafe behavior by CoT monitors and activation probes, and increases the prevalence of consequential reasoning. These results suggest that monitorability should be treated as a trainable attribute of the policy. A dedicated monitorability post-training stage can extend current practices that protect CoT from direct optimization pressure.

Abstract: Every prompt sent to a hosted language model is readable by whoever runs it. We study whether the prompt can remain on the client while the host receives transformed vectors that stand in for it. A language model converts text to vectors exactly once, with a public lookup table; we perform that step locally and pass the vectors through a small trained network (55.6M parameters) before they leave the perimeter. The host runs a frozen 27B model (Qwen3.8-27B) on the transformed vectors and never receives text or token ids. We train the transform against the frozen model as a fixed judge, combining a forward-KL utility term with an auxiliary term that pushes each vector away from its original direction. On 1,490 MMLU questions held out from training, the transform costs 1.5 points of accuracy (80.7% vs. 82.3%), breaking 68 of the 1,226 questions the unmodified model answers correctly (5.5%), while 30.1% of content tokens no longer decode to their original token under a nearest-neighbour lookup.

Abstract: Enterprise adoption of LLMs in agentic workflows is gated by evaluations that span a wide arc of abilities: CRM operations, heterogeneous-document retrieval, personal data understanding, multi-turn planning, tool use, function calling, long-horizon research, and long-context software engineering. These evaluations live today in separate repositories with divergent runners, dependencies, and scoring conventions, which makes cross-benchmark comparison difficult and obscures what a single aggregate number actually represents. We release EGI-BENCH, a unified LLM agent benchmark suite that consolidates eight enterprise facing agent benchmarks (CRMARENA, HERB, PERSONABENCH, USERBENCH, MCP-UNIVERSE, MCPEVAL, BENCHFORCE, LOCOBENCH) behind a single command-line interface with two runtime-budgeted execution modes. On top of the suite we define the Enterprise General Intelligence (EGI) Score, a six-dimensional framework (M1-M6) whose buckets are populated by fine-grained sub-metrics routed from each benchmark rather than by benchmark-level averages. We describe the suite, the sub-metric routing, and the quick/deep budgets, and report results for fourteen frontier and cost-efficient LLMs spanning the GPT, Claude, and Gemini families under quick mode. Top-line EGI ranges from 39.6 (GPT-4o mini) to 65.6 (gpt-5.5), and no single model dominates every capability axis: gpt-5.5 leads task outcome, planning, and retrieval grounding; Claude Sonnet 4.5 leads tool actioning and safety/abstention; and Claude Haiku 4.5 leads engineering quality. Code is available at https://github.com/SalesforceAIResearch/egi-bench.

Abstract: An evidence-review system must allocate additional model judgments and control their use. We study these software decisions on 6,164 frozen manufacturing source-task pairs. Benefit-targeted routing learns recovered agreement minus newly introduced disagreement. On the original 300-case test, 60 consultations yield 202/231 agreements with two human references versus 189/218 for a logistic-risk router. An additional 300 cases were independently annotated twice using equal frozen evidence. Without retraining the historical policies, the corresponding counts are 164/190 versus 163/193; the differences do not resolve an advantage. Strict scalar-Ridge controls share preprocessing, regularization and fitting, and remain inconclusive. Quotation and requested-state replacement gates retain rejected cases and budget costs; their joint gate rejects one original and two additional consultations without changing primary outcomes. Evidence ablations, fusion and bounded updates supply further controls. A completed 200-pair identity audit yields 97.45 percent weighted merge precision after eight author adjudications. This audits a provisional map, while prospective company-disjoint performance remains untested.

Abstract: Medical data remain fragmented across institutional silos because of privacy, governance, and ownership constraints. Federated learning enables medical institutions to adapt large language models (LLMs) collaboratively without transferring raw records. However, independently parameterized low-rank adapters incur repeated parameter and communication costs across Transformer projections, while uniform aggregation may limit adaptation to heterogeneous institutional data. We propose PFedQFT, a quantum-inspired framework for personalized federated fine-tuning with a compact shared adapter. Specifically, PFedQFT constructs reusable orthonormal adaptation bases from compact skew-symmetric generators using a truncated matrix-exponential approximation followed by QR orthonormalization. Compatible projections reuse these bases, while a globally shared full core preserves the flexibility of each projection. Each client further maintains a lightweight private residual operator, whose directions are learned globally and whose coefficients remain local. Experiments on seven public medical tasks and multiple LLM backbones show substantial reductions in trainable parameters and per-round adapter upload volume, with competitive predictive performance on most tasks.

Abstract: Consistency-based defenses for retrieval-augmented generation (RAG) rest on an assumption that a sparse adversary breaks: a poisoning campaign may modify a negligible fraction of the corpus, yet retrieval concentrates its passages into the reader's context. Across four Natural Questions retrieval seeds, at least one injection reaches the top-5 in 94.1% of targeted query--seed rows, and all three form a retrieved majority in 52.9%. Under our short-answer protocol, RobustRAG and ReliabilityRAG reach exposure-conditional attack success 0.512, against 0.576 for undefended RAG, while lowering clean accuracy from 0.328 to 0.216. Coordination, however, leaves a fingerprint: a dense, query-anchored component in the evidence embedding geometry. CoordGuard audits this structure from robustly normalized query-weighted affinities, without poison labels, correct answers, or reader gradients. Its always-filter mode suppresses the selected component whenever one forms; a benign-calibrated gate suppresses it only at unusually high scores and traces a robustness--utility frontier. Pooling four retrieval seeds, e5 always-filtering reduces $\mathrm{ASR}_{\mathrm{exp}}$ to 0.167 overall and 0.138 on parametrically unknown queries, at a 13.6-point clean cost, while a Contriever gate retains clean accuracy 0.315 at $\mathrm{ASR}_{\mathrm{exp}}=0.331$. The reduction transfers across two readers, three tasks, and multiple retriever--auditor pairs, and count-matched controls attribute it to component-targeted removal rather than deletion volume. These findings recast poison detection from judging truth to auditing the structure of retrieved evidence.

Abstract: As LLM-based agents are deployed with increasing autonomy in real-world environments, a critical failure mode emerges: agentic misalignment, where agents choose harmful actions over failure. However, existing safety benchmarks primarily evaluate whether agents comply with explicitly harmful instructions, leaving the conditions that cause agents to choose harmful actions poorly understood. To address this limitation, we propose AM-Bench, a unified taxonomy and 35 scenario evaluation benchmark spanning seven industries for agentic misalignment grounded in the fraud triangle, a well-established criminological framework explaining goal-directed misconduct. We adapt this framework to agentic environments and construct our evaluation suite and detection systems along three dimensions: perceived threat, perceived opportunity, and reasoning. We operationalize each dimension into structured subcategories rooted in prior theory: five threat categories derived from the instrumental convergence thesis, seven opportunity categories drawn from the Emergent Strategic Reasoning Risks (ESRR) taxonomy, and a novel four-question binary reasoning framework synthesized from prior literature. Our evaluation of 10 frontier models reveals a 62-percentage-point gap in misalignment across scenarios, with Gemini 3 Flash and Gemini 3.1 Pro exhibiting the highest levels of agentic misalignment while Claude Sonnet 4.6 displays near-zero rates across all conditions. Our reasoning framework further reveals that models may exhibit harmful behaviours without explicit situational awareness or explanation, complicating both mitigation and detection. AM-Bench scenarios and code are available at https://github.com/AI4Collaboration/Agentic-Misalignment-Bench

Abstract: AI agents increasingly use files and software tools to produce professional deliverables. Such work draws on materials, decisions, and feedback accumulated across assignments. We introduce OfficeTown, a generative simulation framework for professional organizations in which tasks and their context arise from continuing organizational work. From a natural-language description of an organization's purpose and activities, OfficeTown initializes an organization whose members coordinate work, delegate bounded assignments to assistant agents, and review their deliverables. Artifacts, decisions, and unresolved matters persist and shape subsequent work, while responsibilities and permissions determine what each member can see and do. We curate work requests, their context, and review feedback into executable resources for evaluation and learning. We construct OfficeTown Bench with 406 tasks and 4,836 evaluation criteria from 23 simulated organizations across eight professional domains, incorporating domain-expert feedback. Frontier-model evaluations reveal substantial remaining headroom. Supervised fine-tuning and on-policy distillation on OfficeTown data substantially improve Qwen3.5-9B on external professional-work benchmarks, including GDPval. Supplementary evaluations show that the principal--assistant harness achieves higher scores than the single- and multi-agent baselines tested. These results highlight the value of organizational simulation in generating work-grounded data and environments for the continued development of generalist agents.

Abstract: Agentic large language models have been documented to lose reasoning capability when placed inside a harness, the software that mediates their observations and actions. To reconcile these losses, we introduce AutoOnboarder, an autonomous system that adapts an agent to a task the way a trainer onboards a new employee. It places the agent under the supervision of an intelligent ``Onboarder'' coding agent, which iteratively tailors its foundation model, prompt, and harness source code to simultaneously improve performance and decrease cost. Like a real onboarding process, the Onboarder is only given tens of examples, and all evaluation data is isolated on a separate virtual machine to prevent leakage. We report the lessons learned from building and testing AutoOnboarder on genomic-sequencing prior authorization form submission. Autonomously onboarded agents are able to match or exceed the performance of a Gemini 3.5 Flash baseline at up to 29$\times$ lower costs. However, these agents remain imperfect, and the Onboarding process varies greatly in terms of final performance. We conclude that future Onboarding systems require a robust environment for the Onboarder, including a structured thought interface, (controlled) internet access, and a wide model search space.

Abstract: LLM agents are becoming increasingly capable of handling daily workspace tasks by interacting with application environments through tool calls. However, as agents read and write sensitive user data across email, documents, and other applications, a central privacy question is whether they preserve contextual integrity (CI), that is, whether the information flows they produce comply with context-specific disclosure constraints. Existing work either focuses on model-level privacy leakage or evaluates agents in limited application environments with manually constructed disclosure constraints. In this work, we introduce AgentCIScope, a unified platform and benchmark for evaluating agent-level contextual integrity through executable, multi-step tasks across diverse, realistic application environments. AgentCIScope transforms public and synthetic privacy cases into executable agent tasks, with disclosure constraints grounded in three criteria: governing privacy policies, persona-based agreement, and a predefined sensitive-attribute taxonomy. Based on AgentCIScope, we evaluate 12 agents on 936 tasks and find that strong task-completion capability does not guarantee contextual-integrity preservation: single-run leakage rates range from 32.5% to 94.4%, while repeating each task five times further increases leakage by 4-19 percentage points. Leakage varies substantially across attributes: personal preferences and booking details have average leakage rates of 84.7% and 76.5%, compared with 17.1% for credit-card information and 29.4% for SSNs; environments: calendar has the highest average leakage rate as an information source, while WhatsApp ranks highest as an information destination; and governing privacy policies. AgentCIScope provides deep insights into agent privacy leakage and serves as a unified testbed for developing privacy-preserving agents.

Abstract: Rubric-based reinforcement learning (RL) evaluates responses against explicit criteria to construct rewards, offering a viable way to train large language models on open-ended tasks. However, standard methods compress criterion-level feedback into scalar rewards and assign uniform token advantages, leaving structured rubric supervision underutilized. Recent work seeks finer-grained guidance by treating rubrics as privileged information to derive self-distillation targets, but the resulting learning signal is not grounded in criterion outcomes and can lead to reward degradation. This motivates us to reframe the problem as criterion-aware credit assignment, where privileged rubric information serves as a relevance cue for credit allocation rather than a source of imitation targets. Accordingly, we propose CoCA, which derives criterion-level credits from group-relative outcomes and allocates them across tokens based on criterion-induced distribution shifts, yielding both outcome-grounded and fine-grained learning signals. Empirically, CoCA achieves an average gain of +4.2% over standard rubric-based RL across medicine and science benchmarks, demonstrating more effective learning on open-ended tasks.

Abstract: Multi-agent LLM systems (MALS) introduce a privacy-leakage problem that conventional sensitivity-based defenses miss, since information disclosed to one authorized agent can propagate through routine coordination to an unauthorized one, without compromise or malicious behavior. This motivates topology-aware enforcement that blocks unauthorized propagation while preserving authorized disclosure. We evaluate 485 scenarios across coordinator-worker, peer-to-peer, and hierarchical topologies, instrumenting seven communication channels and measuring surface leakage and disclosure to topology-introduced recipients. We introduce RECONS, a Detect -> Decide -> Transform sanitizer conditioning intervention on recipient and execution context rather than sensitivity alone. Topology changes where protected information propagates, with disclosure to topology-introduced recipients rising from 0% to 52.2% while surface leakage stays nearly flat at 60.6-62.8%. This effect persists across four generator models and a second orchestration framework, though magnitude and relay paths differ; framework choice separately drives a 150x leakage spread, so topology alone is not sufficient. RECONS reduces residual disclosure from 96.7% without sanitization to 29.25%, preserving 94.1% of authorized disclosures versus 20.5% under blanket redaction. Detection coverage is the dominant bottleneck, accounting for 95.5% of residual disclosures. These results show that privacy enforcement in MALS must reason about the recipient and communication path realized at runtime, not sensitivity or topology assumptions alone.

Abstract: With ever increasing personalized large language model (LLM) services, the next question to answer is how to improve the quality of services/users' experiences, which is indicated by personalized LLM generation accuracy/quality and inference latency. Although state-of-the-art designs, e.g., parameter-efficient fine-tuning (PEFT), retrieval-augmented generation (RAG) or Full PEFT+RAG, can help, they largely ignore the heterogeneous semantic concentrations of users' historical data and cannot fully unleash the potential of personal data for personalized LLM performance enhancement. In this paper, we propose PEFT Union RAG (PUR), a novel ``fast and accurate'' approach that exploits the semantic concentration of users' personal historical data to split workload between PEFT and RAG, aiming to improve personalized LLM performance in terms of better generation accuracy/quality and lower inference latency. Briefly, PUR partitions user's personal historical data and assigns semantic concentrated ``core'' information to PEFT for offline parametric fine-tuning, while retaining dispersed information for RAG. During the online LLM inference, PUR directly generates responses for ``core'' queries covered by the PEFT region and invokes RAG only for sporadic ``peripheral'' queries. Experimental results show that, by appropriately splitting workload between PEFT and RAG, PUR consistently achieves the best generation accuracy/quality for various users across multiple tasks, while reducing inference latency by 14.4%--46.6% compared with the best performed method.

Abstract: Enterprise AI agents can execute altered versions of approved actions, concurrently reuse the same permission, or retry operations whose responses were lost. These failures arise when approval and resource execution lack a shared contract. KedgeFlow specifies language-independent contracts for trusted controllers to authorize state-changing operations and privileged gateways to enforce that authority independently of agent reasoning. Inspired by OpenFlow, it connects policy checks supported by evidence, signed permission for an exact action, a shared allowance that can be consumed only once, target-side conditions for applying the change, and retained outcome records. Under trusted issuance of exact-action authority, two conditional properties establish at most one successful claim per allowance and link every governed dispatch to a matching grant, claim, and subsequent dispatch authorization, even when the agent is compromised. A reference implementation exposes this execution path to agent systems through the Model Context Protocol (MCP). In the single-resource evaluation, 6,560 concurrent ledger claim attempts yielded exactly one successful claim for each of 82 allowances. Boundary tests rejected unauthorized and altered requests before consumption. A 32-call gateway race produced one committed mock effect, and a retry after a lost response produced no additional dispatch. Deliberately permissive policy nevertheless authorized an undesirable effect, exposing the limits of execution authorization. Evaluation uses three members of the etcd coordinated key-value store on one host, synthetic authentication and authorization inputs, and a mock target that rejects stale requests. Distributed reservations, enforcement of every production access route, native crash recovery, exactly-once effects at real targets, and interoperability between independent implementations remain untested.

Abstract: Multi-agent LLM systems negotiating with a stateful counterpart waste model calls in three ways: polite loops that never meet the counterpart's hidden acceptance condition, malformed outputs that trigger retries, and compliance deadlocks in which the counterpart demands something the agent must refuse. We study a three-part control stack - a 5-Pillar runtime constitution, a 4-tier swarm (Director, three-agent majority vote, Monitor, schema hard gate) and Cognitive Annealing (deterministic deadlock detection, atomic purge of the agent-side context, a canonical recovery message) - against a released adversarial Gatekeeper whose acceptance rules are fixed regular expressions and whose LLM only renders reply text. The testbed has a known solution: it measures whether the stack executes a constitution-aligned strategy against swarm drift and recovers from deadlock, not whether it discovers anything. In five runs per configuration (30 runs; Gemini 2.5 Pro agents, Claude Haiku 4.5 Gatekeeper) we find: (i) the constitution and Director make an acceptable framing possible but not reliable - 0/5 baseline unlocks versus 1/5 and 2/5 with the constitution; when the swarm unlocks it does so in one turn with 7-8 calls and about 15k tokens (67-73% below baseline); when it does not, it costs 17-38% more; (ii) the Monitor and hard gate do not reduce unlocks and leave an audit trail; (iii) under a honeytrap-to-compliance deadlock, LLM-only steering escapes 0 of 5 times while atomic purge plus a canonical strike escapes 5 of 5 (Fisher $p = 0.008$) at the same call budget, with zero calls for the strike. LLM-written strikes failed the deterministic pre-flight 5 of 5 times although an LLM Monitor had approved four. Pre-registered hypotheses on average call and token reduction were not supported. Cost is bounded in every arm by deterministic stop rules; the stack adds recovery at no extra model cost.

Abstract: Enterprise AI agents that share a memory store face two unaddressed risks: sensitive data can leak through legitimately computed results the requester could not derive, and departments can silently compute a same-named key performance indicator (KPI) through conflicting logic. Existing agent-memory systems (e.g., MemGPT, Zep, A-MEM) gate retrieval by content, ownership, and role, not derivation, missing a cached insight that embeds a forbidden column. We introduce the Analytical Memory Unit (AMU), a memory schema that attaches a full derivation (lineage) graph to every cached result, gated by a retrieval policy that serves a hit only when the requester is authorised for every column touched. Provided lineage recording is complete, we prove by construction that the policy blocks retrieval of results derived from a sensitive column outside the requester's permissions, at O(n) worst case -- a conditional design guarantee, not an empirical claim, that excludes derived features encoding sensitive information without naming their source. Eliminating measured leakage required 75-90% recorded lineage completeness, so we treat 90% as a conservative deployment target. Across six experiments, lineage-gated retrieval removes the 18.8-25.5% cross-department leakage naive content-gated memory suffers, keeping 81.5-82.6% of memory reuse at 13.8 microsecond worst-case overhead. A real-agent proof-of-concept with LLM-generated SQL is consistent with the guarantee: zero leaks over 9 round-trips, two conflicts caught automatically -- though a feasibility demonstration, not evidence of production viability. This offers a practical governance layer for shared agent memory, complementing source-layer access control and supporting EU AI Act compliance.

Abstract: Reinforcement learning with verifiable rewards (RLVR) trains a language model on problems that may themselves be confidential, and the trained model can reveal which problems it saw. We study RLVR under prompt-level differential privacy: the released weights must be (ε,δ)-differentially private with respect to the presence of any one training problem. Taking the group of responses to one prompt as the privacy record, our method aggregates their gradients, clips the prompt's contribution once, adds Gaussian noise, and composes the privacy loss across updates, so the budget depends on neither the number of responses per prompt nor the clipping norm; to our knowledge this is the first differential privacy guarantee for RLVR training. We train Qwen2.5-1.5B-Instruct with LoRA at a per-run budget of ε=8 and compare, on the same prompts and at the same budget, a control that removes only the reward signal and two private supervised fine-tuning recipes. The reward signal improves accuracy over the control by 2.65 points on MATH and 3.24 on GSM8K, in every seed; the improvement survives a format-robust scorer, at 1.3 points on MATH, and is not explained by response length. At the same budget the private model outperforms both supervised recipes on MATH and GSM8K by 2.3 to 3.8 points, retains 85--90% of the gain of non-private GRPO on these tasks, and on MATH the noise of an eightfold tighter budget costs at most 1.2 points. The reward effect also carries to CommonsenseQA, an exploratory non-mathematical task. Verifier feedback thus remains a usable learning signal under prompt-level privacy.

Abstract: The last mile toward enterprise AGI is a company that runs itself. Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints. Historical archives are also frequently incomplete and record only what actually happened. They cannot show the outcomes of alternative decisions. We introduce MiniCorp, an office simulator for studying how agents can collectively run a company while generating enterprise data at scale. Using an e-commerce company as a demonstration, MiniCorp connects two interacting worlds. The external world models customers, dynamic competitors, and market mechanisms. The internal world consists of agents that observe events, discuss their options, and make strategic decisions. These decisions have lasting effects on the market, and the resulting feedback informs the firm's later decisions. As the firm and market interact, MiniCorp continuously records the agents' communications and decisions. These records preserve the information available at the time and the business results that followed. Checkpointing allows the same situation to be replayed under different decisions, providing comparisons unavailable in static archives. We evaluate end-to-end fidelity against patterns reported in empirical studies of real markets. These evaluations provide agents with realistic market feedback and reduce the risk that they learn to exploit flaws in the simulator. Our experiments show agents coordinating across roles and adapting their decisions to market feedback. With explicit long-term strategic guidance, they also sustain advertising exploration despite weak early returns. MiniCorp thus provides an environment for studying AI-run companies and a scalable source of longitudinal and counterfactual enterprise data for agent training and evaluation.

Abstract: Indirect prompt injection attacks - malicious instructions embedded in content processed by large language models - remain a major obstacle to safely deploying tool-using agents. CaMeL [Debenedetti et al., 2025] mitigates this threat for an individual agent by separating trusted control flow from untrusted data and enforcing capability-based security policies at runtime. In this work, we investigate whether CaMeL's security guarantees compose in hierarchical multi-agent systems, where agents invoke other agents as tools. We find that CaMeL's guarantees do not compose. We construct a concrete prompt-injection attack that succeeds despite all constituent agents individually operating CaMeL. Our attack exploits the fact that untrusted data can be reinterpreted as trusted input by a downstream agent. We then introduce multi-CaMeL, an agent-to-agent communication protocol that preserves provenance across agent boundaries by separating trusted natural-language instructions from untrusted data passed through a distinct data channel. We evaluate multi-CaMeL's utility on AssetOpsBench and its security-utility tradeoff on MultiAgentDojo, a benchmark we develop by extending AgentDojo to the multi-agent setting. We find that multi-CaMeL reduces attack success rate (ASR) to 0.0%, compared with 0.2% for individual-agent CaMeL and 12.9% with no CaMeL. Multi-CaMeL incurs a utility cost, but this cost trends downward as model capability increases and is modest for the strongest models, suggesting that more capable models better accommodate the constraints imposed by the protocol.

Abstract: Long-horizon agents are now playing an increasingly significant role in assisting humans with complex problem-solving. However, it is exactly their extended interaction history that introduces an underexplored execution-safety concern. Under benign interaction conditions, an agent may execute an action that violates a safety constraint specified many turns earlier. We term this failure mode Governance Hazard from Overlooked Safety Constraints across Turns (GHOST), which may cause irreversible damage. Our experiments reveal that GHOST events are not isolated cases: this failure mode, occurring precisely under benign interaction conditions, yields an occurrence rate of 11.5% on GPT-5.5. Furthermore, we theoretically show that if the residual conditional violation hazard along each safe prefix is bounded below by a non-summable sequence, the execution enters the hazard region almost surely. Leveraging this theoretical insight, we further propose STAR-Guard, a two-layer defense coupling historical semantic safety constraint restoration with pre-execution audit. STAR-Guard restores applicable safety constraints to reduce unsafe proposals, while its deterministic audit layer prevents residual violations from reaching the environment. Consistent with this two-layer design, we observe no GHOST events in our experiments under the GPT-5.5 setup.

Abstract: Enterprise AI agent skills evolve as tool APIs, models, and specifications change, yet final-output evaluation can miss process-level behavioral drift. We present a continuous evaluation framework combining outcome-level and process-level checks, applied to Revenue and Productivity variants of a Business Value Determination skill in an enterprise Value Aware Resiliency system. The framework independently computes per-run ground truth, materializes reusable template tests, and evaluates tool selection, arguments, execution order, and database integrity through programmatic checks and a narrowly scoped LLM judge. We evaluate 240 trials across two skills, two specification variants, two agent harnesses, and three models. Of 175 trials passing all applicable final numerical checks, 162 (92.6 percent; Wilson 95 percent CI: 87.7-95.6 percent) contained another evaluator-detected deviation. Under a broader seven-check final-state definition, 151 of 164 passing runs (92.1 percent; 95 percent CI: 86.9-95.3 percent) still violated a trajectory check. Dependency attribution reduced a mean of 6.34 failed checks per run to 2.65 roots. Specification sensitivity varied by model and harness, with exploratory bootstrap interaction intervals excluding zero for all three Revenue comparisons and one of three Productivity comparisons. Runtime-resolved templates provided reusable regression coverage across the evaluated configurations; longitudinal validation under actual API evolution remains future work.

Abstract: Professional work often starts with a brief request that leaves the professional to work out what is needed, which documents matter, and whether the request's premise holds. We introduce DAYJOB, a benchmark of 130 tasks built by professionals in healthcare (50) and finance (80). The tasks are estimated to take a professional 13.6 hours on average in healthcare and 16.6 in finance. Each task is a containerized Harbor environment with an expert rubric of binary criteria (median 47.5 and 57.5 per task) that an agentic judge applies to the delivered files, and an attempt passes only if it meets every criterion. Across 30 model configurations from 13 developers, the strongest, Claude Opus 5.5, passes 24.7% of healthcare and 23.9% of finance attempts, and the median configuration passes 0.6% and 2.5%. In case studies, agents accept premises that the record contradicts and carry wrong inputs through otherwise consistent analyses. We release all healthcare tasks, 50 of the 80 finance tasks, the evaluation harness, and the leaderboard.

Abstract: Large language model (LLM) agents increasingly operate in private environments and accumulate valuable experience from task execution, tool use, feedback, and local knowledge. Yet such experience is distributed across organizations and cannot be directly shared because of privacy and proprietary constraints. Conventional federated learning is insufficient for this setting, as agent capabilities extend beyond model parameters to memory, tools, rewards, skills, and structured knowledge. In this paper, we formulate \textbf{Federated Agent Optimization (FAO)}, which studies how distributed agents can collaboratively improve through controlled information exchange while keeping raw data, complete trajectories, and private knowledge local. We define FAO as a multi-objective problem balancing agent utility, privacy leakage, and communication cost, and organize its optimization space across policy, memory, tool use, reward, and structured knowledge and skills. We further characterize how private experience can be abstracted, protected, aggregated, and adapted into transferable capabilities, providing a unified view of how agents can benefit from one another without direct experience sharing. Finally, we identify the key challenges of FAO and outline several promising directions for future research toward trustworthy federated agent systems.

Abstract: Enterprise information is represented through artifacts shaped by applications, projects, technologies, organizational boundaries, and local requirements. These structures accumulate over time, creating representational complexity that must be maintained by the enterprise and interpreted by information consumers and AI systems. This paper introduces Enterprise Representation Simplification (ERS) as reducing unnecessary representational complexity while preserving required information within a defined scope, and Enterprise Representation Complexity (ERC), a representation-neutral model for comparing complexity across representation states. ERC characterizes representational extent through four dimensions: Representation Objects, Interactions, Behaviors, and Supporting Sources. Objects, Interactions, and Behaviors form dependent categories, while Supporting Sources characterize representation exposure. ERC is defined at representation and task levels, enabling comparison and distinguishing architectural simplification from retrieval optimization. The paper develops two consequences of ERS. First, representational structures create lifecycle obligations for maintenance, governance, dependencies, change, enhancement, and operation. An economic model distinguishes recurring global representation cost, recurring task-level cost, and one-time transformation cost, enabling evaluation over a defined time horizon. Second, reductions in task-level ERC reduce the representational extent an AI system must identify, relate, and interpret. Text-to-SQL research provides evidence that reduced schema and reasoning complexity can improve reasoning accuracy. ERC is not a universal complexity, performance, or cost metric. It provides measurable architectural variables for comparing representational alternatives, transformation effects, economic outcomes, and AI reasoning performance.

Abstract: Parallel search may generate a correct answer that final-answer voting fails to select. We formulate this consolidation stage as trajectory selection and introduce TRACE (Trajectory Ranking with Aggregated Cross-Rollout Evidence), a lightweight learned selector that ranks completed trajectories using the search evidence behind their answers. TRACE preserves individual query and evidence occurrences, connects rollouts through shared content or document identity, and propagates information across these relations. Each candidate answer then reads the updated states of its own trajectory, preserving retrieval provenance while incorporating evidence from related rollouts. Trained with answer-level supervision over frozen text embeddings, TRACE returns an existing answer without additional search or autoregressive aggregation. One selector per search setting transfers across rollout policies and agent backbones without agent-specific fine-tuning, improving over voting across six WebQA policies and six long-horizon dataset-backbone combinations at $K=16$. On Qwen2.5-14B Base/SFT WebQA pools, TRACE achieves 45.2/49.2% EM, compared with 43.9/48.0% for the strongest Qwen3-32B generative aggregators. On long-horizon FRAMES, GAIA, and BrowseComp, it reaches 78.6% average accuracy, exceeding majority voting by 3.1 percentage points. On Base WebQA pools, TRACE with only 8 rollouts comes within 0.4 points of majority voting over 64. TRACE also achieves at least $10\times$ higher processing throughput than SolAgg, SummAgg, and AggAgent across all seven WebQA benchmarks. These results show that reusing cross-rollout search evidence provides an effective and efficient alternative to heavyweight generative aggregation for parallel search. Code is available at https://github.com/Jaasssoooonnnnn/TRACE.

Abstract: Graph agents extend large language models (LLMs) with the ability to actively explore and reason over knowledge graphs through multi-step interactions with graph tools. However, training capable graph agents typically requires large collections of question-answer pairs and reasoning trajectories, whose manual construction is costly and difficult to scale. Moreover, employing proprietary LLMs to generate such supervision further risks exposing sensitive graph data to external services. Therefore, we propose GraphCert to bootstrap agentic graph reasoning with certified evidence rubrics during post-training. Specifically, the Bootstrapped Graph Quizzer guided by generation controls produces graph-grounded QA pairs and marks supporting evidence, which undergo execution certification and semantic curation. The accepted evidence is then canonicalized into certified evidence rubrics that later reward Graph Solver evidence alignment alongside answer correctness during GRPO training. Experiments on five graph reasoning domains in GRBENCH demonstrate that GraphCert consistently outperforms substantially larger LLM agents and post-training method. Furthermore, our analysis demonstrates that the learned policy transfers robustly across heterogeneous graph domains, suggesting that GraphCert acquires reusable graph-reasoning capabilities rather than domain-specific patterns. These results establish executable self-certification as an effective approach to self-training compact graph reasoning agents. Our code will be made publicly available.

Abstract: Multi-agent LLM systems are increasingly structured like organizations, with roles, critics and shared memory, yet they are evaluated on tasks that last minutes. We ask how such an organization behaves when the ground it stands on keeps moving. Frontier Autolab is a long-horizon testbed in which one simulated firm, voiced by sixteen role personas and a dedicated Red Team, must re-found itself in nine technology eras from 1990 to 2040. Each era is temporally gated: the firm decides from a dated briefing, a historian-judge then reveals what happened and scores the decision on a five-dimension rubric, and lessons enter a persistent Playbook. Six eras are scored against history, one against the live market and two are open forecasts. Across four trajectories (36 era decisions, 180 subscores) we find a consistent foresight-commitment gap: in all 24 historically scored eras the judge rated the firm's recognition of the coming shift above its choice of where to build (mean gap 1.9 points on a 10-point scale), because boards chose the layer their existing assets could reach. Organizational design shaped long-run character. A Red Team armed with numeric kill gates produced fifty years of gated pilots and no product, and the rubric rated this firm highest; firms whose memory stored market-structure lessons pivoted every era, while a firm whose memory stored only validation procedure kept one method throughout. We also show why such results are hard to trust. Scores rise across eras in every run while the judge's own hindsight subscore falls (within-run r = -0.58), so apparent learning is confounded with recall of history, and we trace further distortions to self-judging, briefing selection and score aggregation. We release all records and an API harness, and specify fictional and post-cutoff eras that would turn the testbed into a benchmark.

Abstract: LLM agents operate in persistent collaborative environments involving multiple users, communities, memories, files, and tools. Community boundaries may remain fixed or evolve with changes in membership, roles, composition, and relationships. Agents must complete legitimate tasks and prevent unauthorized disclosure of protected information. Existing evaluations do not fully examine these risks in agent systems. We introduce \textbf{CoSec}, an executable benchmark for evaluating privacy and authorization enforcement in LLM agent systems operating within and across communities. CoSec contains 208 canonical scenarios spanning fixed and evolving boundaries, protected information belonging to the agent owner or other participants, and attacks through dialogue, environmental content, persistent memory, and composed workflows. CoSec executes complete agent systems with persistent sessions, memory, files and tools. It verifies information flows against the active authorization state using execution traces and artifacts. Across harness and model configurations, agents frequently complete benign tasks but violate privacy and authorization boundaries. Privacy behavior varies across harnesses, attack surfaces, and community states, revealing how memory, files, tools, and workflows can carry protected information beyond its authorized scope. These findings show that task utility does not imply privacy or authorization compliance and that authorization in community settings remains an unresolved security challenge for persistent LLM agents.

Abstract: Retrieval-augmented generation systems are extensively instrumented with metrics, benchmarks, traces, and automated judges, but these tools do not decide whether a proposed policy change is safe to release. We present RAGWarrant, an open-source promotion-control framework that treats deployment as a constrained evidence decision rather than a leaderboard choice. RAGWarrant normalizes evaluator outputs and operational telemetry, applies predeclared quality and hard-risk gates, assigns evidence-class claim ceilings, preserves negative outcomes, and emits auditable PROMOTE, BLOCK, REJECT, or INCONCLUSIVE decisions. We evaluate the framework across T2-RAGBench, MultiHop-RAG, CRAG, HotpotQA, synthetic reproduction, and bounded local generative experiments. On HotpotQA, operational savings were blocked because answer quality fell beyond the declared margin. A bounded CRAG study selected a lower-cost quality-tied policy, but related generative gains were unstable and a held-out guardrail failed closed. We claim an auditable promotion-control abstraction, not optimizer superiority, human validation, or production readiness. The tagged artifact reproduces from a fresh clone, runs as a hardened Docker job, accepts external evaluator exports, and verifies artifact integrity.

Abstract: Large Language Models (LLMs) enable us to better understand text documents, including PDFs and Word documents. However, LLMs, as well as more modern LLM agents, i.e., those with tool-calling abilities, typically treat such documents as plain text, ignoring the fact that they are often organized hierarchically into sections and subsections. Extracting this structure, while difficult, can improve efficiency and effectiveness for agents (and humans)---since only sections relevant to a given task need to be processed. Unfortunately, prior work on structure extraction provides no formal guarantees on how well the inferred structure matches the true one. Instead, we target a robust and compact variant that is feasible to infer and useful in practice. Robustness ensures that the text under each subsection header is a superset of the text under the same header in the true structure. Compactness seeks to minimize this superset, reducing agentic cost (or human cognitive load). We propose SHED, a two-stage workflow for inferring a robust and compact structure. The first stage is pluggable with an infinite family of approaches, each guaranteeing robustness for a specific document class. We theoretically characterize the document space using these classes and their hierarchical relationships. Empirically, SHED improves F-1 scores (measuring the robustness--compactness trade-off) by 13%--68% over non-LLM baselines and 9%--15% over expensive LLM-based approaches. Finally, we show how SHED-inferred structures are valuable for agentic document analysis: agents using SHED outperform baselines, achieving 3%--23% higher accuracy while being up to 10x cheaper.

Abstract: Running large language model (LLM) agents in production gets expensive fast. A frontier model (the largest, most capable tier) is accurate but can cost 25 times what a small model costs per token, and the gap compounds once a workflow chains several calls together. Planner-as-Router (PaR) attacks this from a different angle. Instead of leaving model-tier selection to some component downstream, it folds the choice into planning itself. As the planner breaks a query into subtasks, it also assigns each one a model size tier (small, mid, or frontier, ordered by capability and price), so the dependencies between subtasks are visible before any specialist runs. Unlike per-call routers such as cascade routing, which look at one node at a time, PaR sees the whole workflow up front and needs no separate router model or training data. We evaluate PaR with EntBench, a benchmark of 54 enterprise agentic tasks across seven classes, graded by actually running the generated Structured Query Language (SQL) and MongoDB queries against live databases. Over 1,157 evaluations spanning eight routers and three seeds, PaR stays on the observed cost-accuracy frontier. It matches a sink-frontier heuristic (frontier model on terminal nodes only) in accuracy at comparable cost and a faithful FrugalGPT cascade at lower cost, and cuts cost 44% against all-frontier routing while giving up 2.9 points of accuracy. Several accuracy gaps fall inside the plus-or-minus six-point confidence interval of a 54-task study, so we frame PaR's advantage as frontier position rather than a clean accuracy win. We also report a preliminary observation, not a validated result: a small pilot hints that cheap routing may carry a hidden compounding penalty on compositional workflows, which we frame as a hypothesis for future measurement. PaR, EntBench, and all evaluation code are open source.

Abstract: Large Language Model (LLM) agents struggle to scale safely when exposed to vast enterprise toolsets. Providing an agent with access to every internal tool leads to oversized context windows, degraded tool selection, and severe governance vulnerabilities - as system policies defined purely in prompts remain probabilistic advice rather than hard constraints. Existing mitigations, such as multi-agent domain delegation, decentralize audit logs and fail to guarantee policy compliance across sessions. We introduce skilder, a framework that packages capabilities into roles: bundles of skills, tools, and instructions, together with the limits that bound them. An agent begins with a minimal role catalog, learns the roles a task requires, and receives each role's skills, instructions, and tools through a single MCP server. Because tools reach the agent only inside learned skills, the same server enforces the scope of what was learned deterministically. We evaluate skilder against flat-context tool selection and multi-agent orchestration across 13 tasks using six models (10 runs each). Our results show that, when models completed discovery and issued a governed call, the skilder simulated authorization layer enforced governance boundaries: no unauthorized tool call or parameter violation (e.g., a spending-limit breach) executed. Aggregate task pass rates also reflect whether each model followed the discovery protocol and satisfied response-quality checks; those misses are not authorization failures. Furthermore, by allowing agents to dynamically acquire cross-role capabilities mid-task, skilder preserves problem-solving flexibility while providing hard system-level enforcement.

Abstract: Deploying large language models (LLMs) on sensitive data via Retrieval-Augmented Generation (RAG) introduces severe privacy risks. Recent studies apply Differential Privacy (DP) to LLMs with RAG for formal privacy guarantees. However, existing DP-RAG frameworks rapidly exhaust the privacy budget. Although recent efforts attempt to save the budget by narrowing the retrieval scope or sparsifying private generation, these methods themselves cumulatively consume the budget, whereas they could actually rely merely on public information or at a negligible one-time privacy cost. This mismatch fails to align budget expenditure with the model's actual reliance on private data, causing substantial waste on operations that require no private access. To address this, we propose SparsePay-RAG, adopting "only pay what you must spend" as its core principle. Using public information as a zero-privacy prior, it charges the privacy budget only for the private increment. Specifically, SparsePay-RAG narrows the retrieval scope via public topic-guided clustering, adaptively controls private access frequency without privacy cost through isotonic cross-layer trajectory fitting, and compresses per-access budget via DP contrastive decoding. Under strong privacy constraints, experiments show SparsePay-RAG achieves superior privacy-utility trade-offs over baselines.

Abstract: Multi-passage corpus poisoning often repeats one target claim across similar documents, creating correlated lexical and semantic patterns that similarity- and conflict-aware defenses can suppress jointly. We introduce DnD (Divide and Doubt), a targeted attack based on two principles: distributing support for the target answer across stylistically diverse passages, and including a passage that casts doubt on evidence for the reference answer. The first disperses poison-passage representations in embedding space, while the second strengthens target adoption when multiple poisoned passages are retrieved. We evaluate DnD on two open-domain QA datasets across three LLMs and nine RAG configurations, under both black-box and white-box access to the retriever. Across these settings, DnD matches or outperforms prior attacks in most configurations, with its largest gains against clustering- and conflict-aware defenses.

Abstract: No single Large Language Model (LLM) is uniformly reliable across queries, motivating multi-model inference systems that either route among models or combine their outputs. However, routing stops after selecting an initial model, while dense collaboration invokes peers on every query. We show that collaboration is non-monotonic: peers can recover failures that no model solves alone, but can also corrupt initially correct answers. We introduce COMED (Controlled Model Escalation for Multi-LLM Deliberation), a post-anchor controller for selective cross-model collaboration. COMED uses anchor self-consistency, router margin, and a lightweight peer probe to accept confident answers, verify ambiguous cases, and escalate only when collaboration is likely beneficial. We formalize this trade-off with a rescue-harm decomposition showing that selective collaboration improves when rescued errors outweigh collaboration-induced harms. Across medical, scientific, and general reasoning benchmarks, COMED improves fixed and routed anchors in all 16 open-weight settings, with gains up to +10.7 percentage points on MedQA while invoking fewer models and using fewer decoded tokens than dense collaboration. On HLE with frontier models, COMED improves GPT-5.5 from 23.1% to 28.1%, outperforming dense collaboration and achieving the best results.

Abstract: This thesis proposes a framework for "responsible intelligence" to address AI's critical challenges in safety, ethics, and cultural sensitivity. It advances three core areas: First, it improves domain adaptation in specialized fields using active learning and graph-based knowledge to reduce hallucinations. Second, it enhances ethical rigor via a novel decoding-time alignment mechanism that proactively blocks harmful text generation in real-time. Finally, it ensures cultural and multilingual safety through language-specific steering that respects diverse linguistic and social norms. Ultimately, this work provides a blueprint for building next-generation AI that is contextually knowledgeable, ethically sound, and culturally adaptable.

Abstract: Enterprise AI assistants must produce responses that are verifiable and traceable to source evidence. However, retrieval augmented generation (RAG) over heterogeneous enterprise data can suffer from citation drift, unsupported content, and weak source traceability. We present EvidenT (T = Trust + Transparency + Traceability), a lightweight pipeline that verifies extracted evidence against retrieved documents before answer generation, without model retraining. EvidenT combines structured passage extraction with deterministic lexical alignment to filter unsupported content, correct citation drift, and preserve source-span traceability. On approximately 500 real enterprise queries, EvidenT improves gold-source hit rate by an average of 29% over prompting baselines, produces no citations to nonretrieved urls, and achieves near-saturated answer-to-source lexical coverage.

Abstract: The EU AI Act (Regulation 2024/1689) imposes technical obligations on high-risk AI providers, yet Articles 8-15 were drafted for predictive AI and leave seven technical gaps when applied to generative systems, spanning non-deterministic data governance, training-data provenance, continuous conformity, human oversight, open-ended robustness, emergent risk, and generative fairness. We deliver Governance-as-Code (GaC), a framework of 43 machine-checkable acceptance criteria across six compliance modules that run in a CI/CD pipeline and emit Article-indexed audit evidence, and we show the actual Rego policy code rather than merely describing it. Our central commitment is that the Act's open-textured standards ("appropriate levels," "possible biases") become declared, auditable numbers: robustness thresholds are derived from the provider's documented baseline and a state-of-the-art floor, and framing bias is collapsed into eight measurable proxies tested by counterfactual demographic probing. We also correct who owes what, since under Article 25 and Chapter V a downstream deployer relies on the upstream provider's Article 53 training-data summary and documents only the layers it controls, so GaC verifies that summary rather than demanding per-sample documentation the deployer never had. We validate on two enterprise deployments, a high-risk advisory chatbot and a limited-risk content generator, benchmarking against a manual expert audit rather than documentation artifacts that were never designed to enforce compliance. GaC reproduces all of the manual audit's findings, including three penalty-triggering violations, while cutting audit labor by roughly 75%.

Abstract: Agentic large language model (LLM) systems rely on long interaction histories to preserve instructions, tool states, intermediate decisions, and unresolved dependencies, but unrestricted context growth increases computational cost and can reduce efficiency. This study evaluates protocol-preserving context trimming as a reliability-constrained approach for multi-step agentic workflows. Five trimming strategies - recency-based, relevance-based, summarization, protocol-aware trimming, and adaptive budget guardrails - were compared across retained-context levels and workflow-complexity classes using task success, protocol adherence, valid tool calls, token savings, latency reduction, cascading failures, and critical context thresholds. Conventional strategies achieved about 60% mean token savings but lower task success (66.6-77.3%) and protocol adherence (85.5-88.6%). Protocol-aware trimming improved task success to 92.2%, while adaptive guardrails achieved 96.0% task success, 96.3% protocol adherence, and 1.0% cascading failure with 56.0% mean token savings. Retained-context budgets of 25% or less increased failure odds 10.92-fold relative to budgets of 50% or more (p < 0.001). Protocol-aware trimming produced 5.24-fold greater odds of successful completion than conventional methods under aggressive budgets, while adaptive guardrails further increased success odds 2.11-fold versus fixed protocol-aware trimming (p < 0.001). Critical context thresholds also increased with workflow complexity. These findings indicate that reliable context reduction depends more on preserving protocol-critical state than on maximizing token removal, and that adaptive guardrails can improve efficiency, scalability, and reliability in long-horizon agentic systems.

Abstract: We present Salesforce Koa, an enterprise language model built by post-training the open-weight Nemotron-3-Super-120B foundation model with reinforcement learning using Group Relative Policy Optimization (GRPO). Salesforce Koa is trained on public and synthetically generated data, with no customer data, to improve tool use and agentic capabilities while preserving strong general-purpose performance. Its distinctive component is a simulation-to-reward pipeline that expands workflow specifications into persona-conditioned multi-turn tasks with task-resolution rewards grounded in successful tool use for data-dependent requests. For enterprise domains, these specifications are written in Agent Script, Salesforce's declarative language for building Agentforce agents; for public tool-use domains, we synthesize the workflow structure directly. The same simulation and grounded-reward machinery drives GRPO across both. Across public tool-use, agentic-reasoning, and enterprise Customer Relationship Management (CRM) benchmarks, Salesforce Koa improves over its open-weight base, with the clearest gains on multi-turn tool use, and surpasses a strong proprietary baseline while remaining below the strongest frontier models. These results show that specification-driven reinforcement learning is a practical path to specializing open-weight foundation models for enterprise agentic tasks.

Abstract: We introduce NCP-ArchPreview, a latent-space language model that pushes autoregressive pretraining beyond standard next-token prediction (NTP). Alongside NTP, the model learns through Next Concept Prediction (NCP) to predict discrete concepts that span multiple tokens, introducing an explicit and more challenging concept-level objective while preserving standard token-level autoregressive generation. NCP-ArchPreview builds a latent space by constructing a product-quantized concept vocabulary directly from its hidden states, and subsequently learns to predict future concepts via a dedicated Concept Module. These predicted concepts are then fed back to the token level to guide subsequent generation, with NTP and NCP trained jointly end-to-end. We scale this architecture to 8.9B parameters and train it on 5.73T tokens from the Dolma-3 dataset, marking the largest demonstration of a latent-space language model to date. Remarkably, by consuming only 51.3% of the total training tokens, NCP-ArchPreview achieves the final pretraining loss of OLMo-3-7B. Following full pretraining, it outperforms OLMo-3-7B by 2.45 points on the downstream macro-average, including a notable 5.99-point gain on GSM8K. Controlled experiments isolate a clear progression of performance gains stemming from both the latent architecture and the NCP objective. Furthermore, utilizing only 85% of the standard computation, NCP-ArchPreview approaches the training loss of a strictly parameter-aligned 8.9B baseline. The learned latent space remains highly valuable after the pretraining stage: updating just the 17M-parameter VQ module yields a novel, lightweight interface for domain adaptation, while a simple injection of concept representations into a DFlash2 drafter improves the mean accepted length by 4.17% with negligible overhead.

Abstract: Graph-based retrieval can improve multi-hop question answering, but existing approaches often incur high query-time costs and produce diffuse, oversized contexts that reduce generation efficiency. We present LiteRAG, a graph-based retrieval method that replaces expensive retrieval-time LLM control with query-conditioned algorithmic exploration and reasoning-chain context construction. On DistComp, a benchmark for multi-hop retrieval over distributed-systems papers, LiteRAG attains the highest overall quality among the evaluated methods (0.798) while reducing per-query latency by over 100$\times$ and cost by over 99% relative to GraphRAG Global and DRIFT. On UltraDomain, it matches LinearRAG on overall quality while using about 14$\times$ fewer tokens. An ablation study indicates that LiteRAG's query-adaptive thresholding and community-aware hub penalization are the main drivers of its token-efficiency gains.

Abstract: Large language model (LLM) agents increasingly achieve long-horizon tasks by combining foundation models with explicit skills and implicit procedural knowledge acquired through execution. The resulting task-solving capabilities have become valuable proprietary assets, raising a new security question: can a substantially weaker attacker-controlled agent acquire the capabilities of a stronger proprietary agent through limited black-box interaction? Existing skill-stealing attacks recover explicit skill artifacts, yet we show that artifact leakage does not necessarily transfer capability: a weaker agent may possess the same skills but still fail because it lacks procedural behaviors implicitly realized by the stronger agent. Our key insight is that the skill execution gap itself forms a leakage surface, where missing behaviors are exposed through observable differences between successful victim executions and failed attacker executions. Based on this, we present AgentLeak, a black-box capability-cloning attack that identifies capability-critical behaviors from these execution differences and incorporates them into attacker-side skills, while keeping the attacker's model, harness, and tools unchanged. Across 20 task scenarios comprising 600 instances, diverse agent systems, and multiple backbone models, AgentLeak improves task pass rates by over 40% compared with direct skill reuse and recovers more than 80% of the victim--attacker capability gap. Our findings reveal a confidentiality risk in LLM agents: protecting explicit artifacts alone is insufficient, as observable execution behavior can leak the procedural knowledge required to reconstruct proprietary task-solving capabilities in low-capability and attacker-controlled agents.