conversational ai developers

Abstract: Agent memory benchmarks evaluate whether an agent's responses reflect correct knowledge, measuring downstream question-answering accuracy as a proxy for memory quality. We show this evaluation paradigm is incomplete: it conflates memory correctness with retrieval and generation quality, producing both false positives where correct responses mask broken memory and false negatives where incorrect responses blame correct memory for retrieval failures. We introduce FEB-30, a benchmark of 30 fact-evolution trajectories spanning 6 diagnostic categories and 3 difficulty levels, constructed via paradigm inversion: documents are generated from known fact trajectories, yielding deterministic ground truth without human annotation. Each trajectory defines how a fact should evolve across multi-author artifacts (documents, emails, conversations), enabling direct inspection of the memory store after each ingestion step. Applying FEB-30 to a representative memory system, we demonstrate that intrinsic and extrinsic evaluation diverge by an average of 16 percentage points, with 44% of trajectories exceeding meaningful divergence thresholds. Diagnostic decomposition reveals that the system correctly classifies memory operations (create, update, reinforce, supersede) only 20% of the time, a failure entirely invisible to extrinsic evaluation, and that failure modes differ systematically by category and difficulty. We release the benchmark, document generation pipeline, and scoring framework to support reproducible evaluation of agent memory systems.

Abstract: LLM-powered agents increasingly support open-ended activities, yet orchestrating when and how agents interact with humans remains challenging. Existing approaches often provide interaction mechanisms tied to particular tasks or execution patterns, lacking a unified architecture for diverse and evolving interaction needs. We introduce Inflak, a unified four-layer architecture for human-agent interaction orchestration that separates state reconstruction and routing, interaction and execution planning, interaction design, and interface rendering. Its layered design with progressive resolution enables general applicability, runtime flexibility, module composability, and orchestration maintainability. We implement Inflak as agent-native, skill-based plugins within a general-purpose agent framework and evaluate it through a gallery of various application scenarios and interaction paradigms, technical performance testing, and a comparative implementation case study. The evaluations show that Inflak can serve as a general and extensible foundation for orchestrating diverse human-agent interaction flows.

Abstract: Human face-to-face conversation requires coordinating speech, gestures and expressions across both speaking and listening. We present ELLSA 2, an end-to-end full-duplex dialogue model that jointly optimizes speech interaction and streaming gesture-expression generation. An autoregressive, chunk-wise synthesizer combines pretrained acoustic features with distinct listening and speaking hidden states from the dialogue backbone and speech synthesizer, enabling continuous non-verbal generation even when the model is verbally silent. To translate dialogue understanding into vivid semantic gestures, we introduce contextual semantic calls and template guidance, separating whether to invoke an available semantic gesture from how to realize it. We also establish a curated data and evaluation pipeline for dialog-aware co-speech gesture generation and investigate reference-based multimodal large language models (MLLMs) as automatic judges. Experiments show that ELLSA 2 remains highly competitive with open-source full-duplex speech interaction models and outperforms cascaded baselines across dialog-aware gesture and expression metrics. Semantic template guidance further significantly improves semantic gesture quality. Strong system-level correlations with human judgments support these reference-based scores as useful preliminary metrics for full-duplex and semantic gesture assessment. Together, these results advance dialogue modeling from spoken responses toward coordinated multimodal interaction.

Abstract: Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents. However, existing memory methods often optimize retrieval and memory construction separately, without a shared reward space for evaluating search, writing, and stopping decisions. To address this, we propose All-in-One Memory (AiO-Mem), which organizes these operations as transitions over a shared evidence state and directly evaluates the semantic quality of each action in a unified reward space. AiO-Mem compiles heterogeneous histories into source-traceable multi-view events, providing a shared substrate for retrieval and memory writing. A unified memory policy actor then rolls out complete interaction trajectories over this shared evidence state. A separate trajectory-conditioned process reward model is trained through evidence-grounded supervised fine-tuning and trajectory-level DPO to assign process scores to the collected trajectories. These scores are aggregated into an action-balanced trajectory reward and used by Process-Aligned DAPO (PA-DAPO) to optimize the memory policy. Across long-context memory benchmarks, AiO-Mem consistently improves answer accuracy over both static memory systems and RL-trained memory agents.

Abstract: Backchannel prediction has been studied almost entirely in dyadic conversation. We introduce a multi-party benchmark based on the AMI corpus, comprising 682 masked-listener views from 171 meetings, 190 speakers, and 18,697 backchannel events, with a person-disjoint held-out split. A state-of-the-art dyadic model applied zero-shot to meeting audio performs at chance (AUROC 0.499); nevertheless, its frozen acoustic features remain informative: a linear probe reaches 0.704, and retraining the predictor raises performance to 0.751. Retraining reveals a second limitation. Listener conditioning improves prediction for listeners seen during training but not for unseen listeners, and the gap remains under capacity reduction, listener-adversarial training, per-listener adaptation, and oracle lexical conditioning. Adversarial training removes only part of the speaker-identity information, while stronger removal hurts prediction, suggesting that identity is entangled with cues that are useful for backchanneling. A within-model control helps explain this pattern: with the same features and data splits, turn-onset prediction transfers to unseen listeners, while backchannel prediction does not. Backchannel rates also vary about twice as much across individuals as turn-onset rates. Since backchannels occupy only about 1% of frames, frame-level F1 is strongly affected by the base rate. We therefore report AUROC alongside event-F1 on listener-active regions. We release the benchmark and evaluation tools at https://github.com/HafsatiMohammed/bc_multiparty_release.

Abstract: We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continuing to exhibit capabilities on benchmarks that exceed those implied by the assigned role. For example, when a model is prompted to assume the role of a kindergarten student, one might expect its performance on a mathematics benchmark to reflect kindergarten-level ability rather than expert-level proficiency in solving calculus problems. We introduce RoleCapBench, a curriculum-grounded benchmark for evaluating RCL across six educational roles and four assessment levels spanning elementary school through A-level, and use it to evaluate three open-weight reasoning models. We find that although the models can generate stylistically convincing in-role responses, they consistently fail to align their underlying capabilities with their assigned roles. Naive role prompting yields strong role-voice scores of 1.218--1.389 while retaining above-role accuracy of 0.811--0.898. RCL persists across a range of prompting conditions, including prompts that explicitly instruct the model to match the role's capability level. To mitigate this problem, we propose Injection, an inference-time intervention that combines explicit, role-specific capability guidelines with a guiding prefilled response prefix. Injection improves role-capability alignment across models, reducing above-role accuracy by up to 0.562 while preserving in-role accuracy with a marginal drop of less than 0.058 across most models. All artifacts, including scripts and evaluation data, will be released upon acceptance.

Abstract: Adaptive RAG uses signals such as confidence, relevance, support, and retrieval quality to decide when to search or correct evidence. In multi-step retrieval, however, these local signals must be combined into a persistent view of what the current evidence supports, what remains missing, and which action should follow. Existing methods often use such signals as separate triggers, making it difficult to preserve a coherent evidence state across a trajectory; we call this problem evidence-state fragmentation. We introduce BELIEFRAG, a closed-loop controller that updates an explicit state over sufficiency, reliability, conflict, uncertainty, evidence gaps, and acquisition cost, then chooses among retrieval, query rewriting, verification, answering, stopping, and abstention. Across six QA benchmarks with gpt-oss-120b, BELIEFRAG reaches mean token F1 0.572 with 3.89k tokens per question, outperforming fixed iterative retrieval (0.555 F1) while using 39% fewer tokens. The same quality-cost pattern transfers to Qwen3-32B, where BELIEFRAG reaches 0.552 F1 versus 0.523 for iterative retrieval while using 35% fewer tokens. Analysis shows that the main gains come from corrective re-retrieval rather than pruning alone, while several belief dimensions are redundant and calibrated answerability plays the strongest operational role. Calibration improves threshold stability across related evidence sources, although source shift can still invalidate the same decision signal.

Abstract: Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-agent systems, Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-agent collaboration leaves this loop open: collaboration is pre-defined at the operator level, topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned Agent Collaboration: a trainable Collab-Director constructs teams of complete Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.

Abstract: Modern LLM agents increasingly tackle complex tasks through interactive, long-horizon exchanges with users, while existing benchmarks generally assume that users always accurately and sufficiently communicate a fixed intent. However, this oracle communication assumption rarely holds in practice: users may miscommunicate, change their goals, and run out of patience. We define this task setting as Interactive Intent Alignment, where agents must recover and continuously track the user's current intent despite imperfect communication and evolving goals. To study this setting, we introduce Drift-Bench++, a principled benchmark construction pipeline for verified executable tasks with controlled misalignment and intent shifts, along with an interaction protocol featuring finite patience, diverse simulated users, and silent interaction-conditioned shifts. We further develop GRIP, a comprehensive evaluation protocol covering task grounding, user realism, inquiry effectiveness, and adaptation to evolving intent. Across diverse environments, models, and interaction conditions, stronger interaction consistently helps but remains far from oracle performance; Validation on deployed ProdAgent sessions further shows that the modeled failures are prevalent and consequential in deployment. By providing a unified, executable benchmark for interactive intent alignment, Drift-Bench++ offers a foundation for evaluating and advancing agents under realistic communication and evolving intent.

Abstract: Long-term memory enables language models to use past interactions in future conversations. However, evidence needed to answer a question may be scattered across distant turns, while the question itself omits clues needed to locate it. Retrieved facts can reveal these clues, motivating retrieval decisions conditioned on evidence already found. We introduce MERA (Missing-Evidence Retrieval Augmentation), which separates globally searchable memory from a question-specific evidence state. Verified evidence guides subsequent retrieval without restricting access to the global memory. We train a lightweight planner through reinforcement learning, rewarding queries that recover previously missing evidence. MERA achieves strong answer accuracy across Qwen3-30B and GPT-4o-mini backbones. With Qwen3-30B for evidence processing and answer generation, the trained 0.6B planner achieves 77.40% accuracy on LoCoMo and 71.29% on LongMemEval-S, exceeding a 30B planner without retrieval-grounded training by 4.10% and 3.96%, respectively. On LoCoMo, later retrieval rounds increase cumulative evidence recall from 55.5% to 80.5%.

Abstract: When conversing with large language models (LLMs), users often begin with a simple question and build towards a multi-hop question through follow-up turns. Retrieval-augmented generation (RAG) and its graph-based variant (GraphRAG) have become the dominant approaches for grounding LLM responses in external evidence, yet both are evaluated almost exclusively on single-turn, fully specified queries. We systematically investigate this evaluation mismatch through a large-scale simulation study. Building on prior work on multi-turn LLM evaluation, we transform questions from multi-hop question answering (QA) benchmarks into underspecified conversations and evaluate ten LLM assistants with eight retrieval systems across 1.5 million simulated conversations. Our findings reveal that multi-turn interaction causes widespread performance degradation, incurring relative performance drops of up to 21% and increasing unreliability by 47%, making RAG systems simultaneously less accurate and less reliable. We identify two distinct failure modes behind this degradation. Systems are either lost in translation, where conversational rephrasing distorts the retrieval query, or lost in conversation, where retrieval succeeds but the LLM fails to synthesize evidence distributed across turns.

Abstract: A personal language agent that acts for its owner across private and shared conversations can learn a fact from one audience and later place it in the context it assembles for another. We study authorization before context across the whole memory lifecycle. Each memory item carries the audience present when it was recorded; derived items are partitioned by audience, receive the intersection of their sources' audiences, or are suppressed; an audience widens only by an explicit, object-specific grant; and an item enters a model attempt only when every current viewer belongs to one of its authorized audiences, with unresolved viewers failing closed to public-only. Under explicit identity, provenance and complete-mediation assumptions, this admission is sound and policy-complete on the exact assembled context, enforced by exclusion rather than by model behavior. We realize it in two independently persisted reference architectures, a flat store and a relationship graph, and, descriptively, in a native agent-memory runtime. In a prospectively frozen confirmation over 10,000 multi-party histories, no forbidden item entered any architecture's context, whereas unscoped retrieval exposed forbidden items in 82% of its contexts. Entitled recall matched policy-equivalent baselines exactly and exceeded unscoped retrieval by 0.30 Recall@5, with a Holm-confirmed advantage that grows with distractors. No architecture produced a wrong-principal substitution, but unscoped substitutions were too rare to establish the prespecified joint decision.

Abstract: Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to address a partner's constraints and advance their goals without adding words that do not help the interaction. In this paper, we propose Teacher-Assisted Communication Training (TACT) to improve social goal attainment while reducing communication cost, making interactions with agents more productive and less demanding. We first characterize communication efficiency in terms of action strategy and expression, whose effects extend beyond the current utterance to the partner's response and subsequent exchanges. We design TACT to revise student-generated actions, test the revisions through partner responses, and distill useful feedback into the student. An expression specialist removes unnecessary detail while preserving the intended action, while a strategy specialist proposes alternatives that may better address the partner's constraints. To determine which revision helps, TACT samples a partner response for each candidate and selects a teacher reference by balancing local goal support against action-token cost. That reference guides on-policy distillation on the student's own generation prefixes, allowing the student to act independently at deployment. We evaluate TACT on SOTOPIA and AgentSense. On SOTOPIA, it achieves the highest Goal among the evaluated methods on All and Hard while using substantially fewer target tokens than SFT+SDPO. On AgentSense, it improves goal success over the initial student while reducing target tokens and interaction messages.

Abstract: Lookup-based memory has been a promising way to scale the parameters of large language models (LLMs). It retrieves learned representations of local token patterns, such as n-grams, instead of reconstructing them through successive layers of computation. However, existing designs such as Engram treat each retrieved embedding as a monolithic unit. Each embedding is stored in its own hashed slot and modulated by a single scalar gate. As a result, polysemous patterns cannot selectively read out the components of their memory that are relevant to the context. Moreover, parameters are shared only through hash collisions, which are largely unrelated to semantics. We propose FactorEngram, a factorized n-gram memory with basis-level contextual gating. FactorEngram retrieves sparsity-regularized coefficients over a dictionary of basis vectors shared across patterns, so related patterns can reuse common components. The same dictionary is also used for gating. The backbone hidden state is scored against each basis vector to gate the corresponding coefficient before reconstruction, which lets the context modulate each memory component individually. FactorEngram also covers both individual tokens and multi-token n-grams, and we systematically study where the memory branch should be inserted. On 340M- and 1B-parameter Transformer backbones, FactorEngram improves language modeling and downstream task performance. Ablation studies confirm the contribution of each component and identify insertion before the attention sublayer in the middle layers as an effective configuration.

Abstract: What shapes a model-generated inquiry when no discussion question is supplied? We introduce a bottom-up forum framework inspired by Philosophy for Children, in which language-model agents read a philosophical narrative, propose and select questions, and develop a shared conclusion without a privileged model facilitator or aggregator. Across 576 forums, contrasting Aristotelian value personas interacted with a blank-slate participant receiving no value-specific instruction. The blank slate remained neutral and was selected more often for conclusions than questions. Yet inquiry narrowed in both form and source: varied initial questions increasingly became either/or alternatives, while discussion concentrated on directions already explicit in the text. Our ECO framework traces this source focus by distinguishing explicit philosophical framing, characters' modeled inquiry, and open-ended narrative material. Across analyzed chapters, participants drew most often on explicit framing. We describe this as a pedagogical constraint: freedom to formulate questions did not necessarily produce freedom from source framing.

Abstract: On-policy distillation (OPD) trains a student on its own rollouts using dense supervision from a teacher. In multi-turn environments, a mistake at a critical decision step can redirect the subsequent rollout toward poor outcomes. We use low teacher confidence on student actions to select high-uncertainty steps for correction. In a controlled ALFWorld study, a single teacher correction at a low-confidence step improves subsequent student behavior and task success, motivating selective intervention during distillation. We propose UOPD, an uncertainty-aware intervention method for on-policy distillation. At low-uncertainty turns, UOPD executes student actions and applies the standard OPD loss. At high-uncertainty turns, it samples and executes teacher actions and trains the student to imitate them through supervised fine-tuning, which minimizes forward Kullback-Leibler divergence in expectation. UOPD utilizes adaptive uncertainty thresholds to target a scheduled intervention rate. Empirically, we evaluate UOPD across a broad range of agentic tasks, including ALFWorld, WebShop, and Search, demonstrating its superior performance over OPD methods and their variants. UOPD improves WebShop score by up to $15.8\%$ relative to standard OPD.

Abstract: We introduce NSV-Shift, a contrastive benchmark for evaluating whether speech-to-speech models can understand non-speech vocalizations (NSVs) and adapt their responses accordingly. Each pair contains two conversations with identical lexical content that differ only in the NSV embedded in the final turn. Our pilot contains 22 human-verified pairs (44 audio conditions) and evaluates five models on NSV perception, emotion understanding, and response adaptation. Results show that models generally perform better at detecting NSVs than at interpreting their fine-grained emotional meaning or producing appropriately differentiated responses. The data construction pipeline, dataset, and evaluation pipeline are publicly available at https://github.com/ChenzwNina/nsv-construction.

Abstract: Task-oriented dialogue requires maintaining and updating information across turns, yet language models expose no explicit belief-state object. We study how conversational state is represented, updated, and used inside eight instruction-tuned language models from four families on MultiWOZ and SGD. Structure and values separate: which domains, slots, and requests are active is linearly readable just before the model acts, whereas exact values are far more readable where the user stated them. After a user changes a value, both values remain accessible at their mentions, and causal interventions show that both continue to influence the model's action. In natural closed-loop interaction, query failures separate into cases of weak structural support, incorrect value resolution, and failure to deploy otherwise-supported constraints, with targeted interventions producing systematically different repair behavior across these cases. These findings motivate a state-action controller that starts from the base model action and selectively edits it using structural readouts, without requiring a complete predicted belief state as an intermediate representation. On held-out MultiWOZ interaction across five models, it raises the base model exact-query accuracy from .318 to .621 and task success from .272 to .371 at negligible added cost. Overall, reliable interaction requires not only retaining conversational information, but resolving which available constraints currently apply and ensuring that they govern action.

Abstract: While memory enhances personalization in LLM-based conversational agents, it suffers from memory misalignment, where memories violate user expectations. We present a mixed-methods investigation to characterize and mitigate memory misalignment from user perspectives. First, we collected data from memory usage (N=28, 457 entries) and diary study (N=32, 304 reports), which yielded a taxonomy spanning 14 misalignment types across memory intake, storage and management, retrieval and interpretation stages. Second, four co-design workshops with 12 experienced HCI researchers derived a design space to tackle memory misalignment issues, consisting of 12 candidate interaction strategies structured across interaction form, placement and intrusiveness dimensions. Finally, a speed dating with 121 users reveals preference heterogeneity, where users prioritize proactive controls over cognitively demanding causal graph inspections or passive audit logs. Synthesizing these findings, we highlight the tension between supervisory agency and interaction overhead, and advocate for friction-aware memories that balance user oversight with conversation smoothness.

Abstract: Existing summarizers for memory systems are typically optimized for human-facing criteria such as faithfulness, which misaligns with their true objective: preserving the evidence needed to support future queries. We show that conditioning summarization on query-answer pairs substantially improves answer quality, and that this utility-aware behavior is transferable across queries. Motivated by these findings, we propose MemSuit, a self-distillation framework in which a teacher summarizer, conditioned on observed query-answer pairs, produces utility-aware memory entries that a student learns to reproduce from the raw conversation alone. To prevent collateral erasure where conditioning on a single query-answer pair discards evidence relevant to other plausible queries, the teacher decomposes each block into multiple self-contained entries that preserve distinct query-relevant facets as independently retrievable units. To align the retriever with the compact, fact-dense style of teacher entries, we further fine-tune the embedding model with a contrastive objective supervised by teacher entries. Across a diverse suite of conversational query types, MemSuit consistently outperforms state-of-the-art baselines, confirming the value of grounding memory in downstream utility.

Abstract: A two-hop question makes a language model retrieve twice: once to produce a bridge entity, and once to retrieve with it. Transformers resolve that entity in their early layers. Hybrid models replace most of the attention with a recurrent state, so where the key becomes usable, and where it is spent on an answer, is not known. Answering "Where is the ball?" from "the ball belongs to Alice" and "Alice is in the garden" turns on Alice, a name the question does not mention. A model could reach garden through Alice, the key it computed, or through where the fact sits in the prompt. Across twelve models, dense and hybrid, we move a hidden state from one story into another where the two routes lead to different places, and read off which place the model gives. An attention layer converts the key in every model we tested, and crossing that layer removes its usable effect, all of it where no attention follows. In sequential hybrids this makes the answer a handoff: recurrent layers carry the key forward, and attention spends it. Retrieval does not always stop there. Writing a different fact into the memory of the recurrent layers that follow the last attention layer can move the answer toward that fact, multiplying its odds by 1.3 to 2.7, and which hybrids do this is not settled by their architecture or training. That read is addressed by the key, not by position: recurrent state in a hybrid is not only a carrier, but a memory that later layers can query.

Abstract: Work on memory for multimodal agents optimizes what is written, updated and retrieved. Between retrieval and the answer, however, is a stage that multimodal memory evaluations do not isolate: what of the retrieved memory reaches the model, and in what form. We call it delivery, and a controlled decomposition on MemLens locates the remaining room there. With the retrieved evidence set exactly fixed, delivering the original pixels instead of withholding them raises accuracy by 13.87 points on an 8B backbone, whereas making retrieval perfect on those same messages improves it by 2.31. Delivery is the larger term on all three MemLens backbones and grows with backbone strength; retrieval grows too, without closing the gap. We propose DeliverMem, an instantiation of delivery as three decisions: keep the original modality, give each item a readable identity, and state when it was seen, with a retrieval-side adapter for the one property delivery cannot supply. Each is measured against a delivery-matched control that alters only its own variable. DeliverMem leads the strongest published memory agent on MemLens at all four context lengths, and beats DMV-Bench's own strongest method at every setting on both backbones. On MemLens it does this on a tenth to a seventieth of the input. Each decision helps only where the question lacks what it supplies, and is null elsewhere. A single fixed configuration nonetheless leads both benchmarks, without training any component or modifying the stored records. Project page: https://avalon-s.github.io/DeliverMem/

Abstract: Long-term memory is essential for multimodal agents that interact with users across sustained conversations. However, user memories are not always explicitly stated: they may also be implied by recurring background objects in images, ambient sounds in audio, or other peripheral multimodal cues. Existing benchmarks largely focus on text-only memory or explicit multimodal evidence, leaving implicit multimodal cues underexplored. We introduce CUE-Mem, a text-image-audio benchmark for evaluating long-term user memory from implicit cues. CUE-Mem contains 2,674 questions across explicit and implicit evidence settings and covers four tasks: Entity Recall, Long Pattern, Personalized Recommendation, and Answer Refusal. Across textualized memory systems, implicit performance remains far below oracle evidence, locating the main bottleneck in preserving and retrieving subtle cues rather than question answerability. Increasing caption detail recovers more of this evidence, but brings uneven gains and rapidly growing token costs, motivating native multimodal access. Yet native access does not uniformly resolve the bottleneck: evidence use depends strongly on the backbone, while multimodal indexing introduces substantial retrieval noise. CUE-Mem provides a testbed for memory systems that selectively retain, retrieve, and use subtle multimodal evidence.

Abstract: Inserting advertisements (ads) into consumer-facing LLM output is emerging as a new business model, but there is little shared evidence on how such ad insertion should be evaluated or how it affects user preferences. We introduce LLMAdBench, a human-preference benchmark for studying advertising in LLM-generated content. The benchmark isolates a simple but practically important decision: given a user conversation, an LLM response, and a matched advertisement, where should the ad be placed? Our dataset compares pairs of responses that differ only in ad position while holding all other conditions fixed including the user query, base answer, advertisement, and disclosure condition. Human annotators evaluate each pair based on six criteria from both advertiser's and user's perspectives. The resulting benchmark contains more than 18000 human judgments across two disclosure conditions: explicitly labeling the ad as sponsored and merging it into the response without disclosure. We use LLMAdBench to evaluate eight frontier LLMs as preference judges and find that they are not reliable substitutes for human evaluation. Even the most stable models reverse roughly one quarter of their decisions when the presentation order is swapped, agreement across models is low, and their placement preferences differ systematically from those of human annotators. Moreover, LLMAdBench contains substantial learnable signal. In particular, a Qwen3-8B model fine-tuned on the human preferences improves substantially over its base model and outperforms all zero-shot frontier judges on the held-out prediction task. Beyond model evaluation, LLMAdBench provides quantitative evidence on the advertiser-user trade-off and shows that the sponsorship disclosure systematically changes users' preference over ad placement.

Abstract: Agent memory grows as agents read inputs, reason, and call tools. Longer histories increase inference cost and eventually exceed the context window. LLM-based summarization reduces this history but adds latency and provides no explicit bound on information loss. We propose LAM, a Lossy Agent Memory system with three components: a deterministic deduplication rule with a substitution bound on retrieval scores - a bound on score perturbation, not a certificate of unchanged ranking; a memory manager that preserves the cached prefix and overlaps compaction with inference; and a performance model that estimates compaction costs before deployment. On 600 agent trajectories, LAM removes 22.47% of observation tokens while retaining 99.984% of the measured gold-patch evidence. At a fixed deletion set, the performance model predicts a 71.4x-91.6x end-to-end speedup from removing records before prefill instead of deleting them from a prefilled context. That benefit comes from the schedule rather than the rule and applies to any prefix-preserving test.

Abstract: We present a framework for evaluating and improving a large-scale, multi-agent shopping assistant in production, and report lessons from its use. Offline evaluation of such a system faces three obstacles. (i) A logged conversation cannot be replayed against a modified system, because a different response changes every turn that follows. (ii) The unchanged system itself varies from run to run. Its LLM components are stochastic, and in product search the available products, their prices, and the customer's personalization signals change. (iii) Aggregate quality scores combine distinct behaviors, so they show that quality has changed but not which behavior caused the change. Our framework addresses each obstacle in turn. For a reported behavior, an Evaluation Harness generates targeted assertions and a fixed cohort of customer scenarios. It then reproduces the behavior in a local instance of the assistant through grounded user simulation. Instead of replaying the log, the simulator writes new customer turns conditioned on the recorded messages and context. Repeated runs of the unchanged system form a stored baseline. An Improvement Orchestrator turns the assertion results into hypotheses, implements each as an isolated modification, and compares it with the baseline using paired percentile bootstrap intervals over scenario-level differences. When an investigation ends, the harness may propose revisions to future evaluations, subject to human approval and without altering past decisions. We report production investigations with this framework. Assertion profiles showed which positions of a product carousel a failure affected, and repeated runs distinguished a real improvement from run-to-run fluctuation. Audits of the evaluation itself found a judge that lacked the evidence it needed and a model setting that was configured but not applied.

Abstract: Existing agent memory frameworks mainly create memory through an agent's interaction with the factual world, e.g., remembering feedback from actions taken to improve performance on future tasks. However, these frameworks seldom ask the "what if" question during memory construction: what if a different action had been taken, would the feedback have changed, and how could this feedback become useful memory? Obtaining such feedback directly in an active environment can be expensive and can alter the state needed for comparison. In this work, we introduce COUNTERMEM, a reinforcement-learning framework for constructing and using verified counterfactual memory across tasks. After a failed action, COUNTERMEM evaluates local alternatives from a copy or reset of the original state using executable world models, such as tests, proof checkers, and solvers. It stores improvements with the original and corrected actions, checked outcomes, and conditions for reuse. A learned memory-use policy selects a retrieved record or skips memory to balance task success and interaction cost, while the base LLM remains fixed. Both memory and policy are frozen during held-out evaluation. We evaluate COUNTERMEM on 12 benchmark settings across six domains. With gpt-oss-120b, COUNTERMEM improves both ReAct and Reflexion on all 12 benchmarks across six domains, averaging a gain of 12.6 percentage points over their unaugmented versions. In the four-domain comparison across two backbones, task-run tokens decrease by 7.7-42.0%, excluding offline selector-training costs. Further analyses show that removing verification or persistent storage weakens the gains, while applying verified corrections to unsuitable decisions can reverse them. Code will be released upon acceptance.

Abstract: "System One" decision models such as TypeSafe's Jev and its open counterpart Laya answer typed questions about a text in a single forward pass with calibrated probabilities, but they cannot ask for missing information: when a first message does not say what separates two departments, they guess. We present LAVOIR (Laya with Value-Of-Information Routing), which places the candidate pieces of missing information (slots) in the input next to the answer options, so that one forward pass returns both the decision distribution and, for every slot, the expected gain in the probability of the correct decision if the user were asked about it. VOI targets need no human labels: gold decisions come from schema rules, an LLM only verbalizes messages and answers, a model from another family checks every text, and pairing each message with several profiles makes regression on realized gains estimate the expected gain. A Gini-impurity cap bounds the predicted value by what a calibrated model can still gain. In a controlled study, decisions on seen schemas are statistically indistinguishable from the Bayes ceiling. The final model's question policy matches a greedy oracle VOI policy on seen schemas (AUC 0.799 vs. 0.797), and with at most 0.5 questions per conversation it is 14.1 points more accurate than never asking. On real ABCD conversations, one real exchange raises accuracy by 8.3 points where LAVOIR asks and leaves it unchanged where it does not; on SGD the cap lowers the asking rate from 93% to 8.6%. On Laya's twelve benchmarks LAVOIR is above Laya's reported scores on seven, and it answers a question in 31 ms (median, GH200).

Abstract: Agent memory systems are increasingly used to maintain long-term user preferences, task states and evolving facts, but current evaluations often collapse memory behavior into final-answer accuracy. We introduce MemProbe, a cognitive-science-inspired framework for diagnosing stability-plasticity tradeoffs in agent memory. The framework is motivated by a core insight from cognitive memory research: memory is reconstructive and shaped by interference, source reliability, reinforcement, and reactivation. MemProbe turns this insight into four reusable experimental paradigms (interference, misinformation, consolidation strength, and reconsolidation window) that manipulate when a memory should be updated, preserved, or treated as uncertain. It further decomposes correctness into behavioral profiles that reveal how systems update, preserve, attribute, and temporally organize information. We instantiate these paradigms in a 56-episode diagnostic suite and evaluate six incremental memory systems under a unified protocol. Results show that systems with similar aggregate scores exhibit distinct behavioral profiles. MemProbe provides such a diagnostic lens, turning aggregate performance into interpretable profiles of memory maintenance over time. Code is available at https://github.com/jq-ding/MemProbe.

Abstract: Despite strong performance on standard benchmarks, it remains unclear whether large language models (LLMs) evaluate social pragmatics in ways that align with human judgments. We evaluate LLM politeness judgments using two English-language datasets with complementary annotation formats: continuous human ratings and three-way categorical labels. Across the seven evaluated models, we find that inter-model agreement is stronger than model--human agreement. Strategy-level analyses suggest that model--human alignment is associated with explicit linguistic cues, while some rapport-building strategies occur more frequently in misaligned cases. In the categorical task, model predictions exhibit systematic neutral compression, characterized by the overproduction of Neutral labels and the underprediction of Impolite labels. This pattern persists when expert consensus is used as the reference on a diagnostic subset. Our findings highlight the need for pragmatic evaluations that go beyond aggregate agreement metrics by examining directional patterns of model--human disagreement across different human references.

Abstract: Small language models (SLMs) are increasingly paired with knowledge graphs (KGs), yet end-to-end KG question answering conflates graph access, search, navigation, reasoning, and answer generation. This coupling makes it difficult both to determine whether an SLM can faithfully execute the reasoning path implied by a question and to attribute failures to navigation rather than to other stages of the pipeline. We isolate this capability by employing the THESEUS navigation and traceability framework and using frozen, off-the-shelf SLMs as local action policies. At each hop, the environment exposes the legal outgoing graph actions, and the model selects one executable graph action and decides whether to stop, without task-specific parameter updates, model-controlled beam search, or free-form answer generation. This controlled setting allows us to evaluate terminal-answer accuracy with Hits@1 together with path fidelity, using Path Edit Distance (PED) as the primary trajectory metric. Across the Kinship and MQuAKE-ST KGQAs, similarly sized local models differ substantially in answer accuracy and path fidelity, with the two metrics sometimes favoring different models. This model-dependent behavior also extends to prompting, as a single demonstrated trajectory can improve or degrade navigation depending on the model. These results motivate evaluating SLM graph reasoning beyond endpoint accuracy alone.

Abstract: Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics.

Abstract: Despite advances in long-context inference, large language models (LLMs) remain fundamentally limited by the key-value (KV) caching mechanisms that are necessary for stable computation. Techniques such as selective token eviction and pruning have vastly mitigated these issues, but often discard core information to manage the growing cache. In this paper, we propose Attention with Routed Memory (ARM) a novel KV caching structure that introduces a fully differentiable, fixed-size memory system organized as a hierarchical router. Via a Gumbel-Softmax, ARM learns to select memory slots and perform sigmoid-gated updates that softly combine new and stored information, avoiding hard eviction and reducing information loss. By further training a policy to dynamically select varying amounts of memory at inference, ARM adapts its accesses for both simple contexts and inputs that require deeper reasoning, enabling more scalable and effective retrieval on both short- and long-contexts. Experimental results on standard commonsense and long-context reasoning benchmarks demonstrate that ARM achieves superior performance and efficiency compared to fixed KV-caching approaches, while remaining efficient and scalable in terms of both memory and generation latency.

Abstract: Conversational AI assistants remember what people tell them, and for disabled people, that often includes disability. We interviewed 12 adults with disabilities in the United States who use LLM-based assistants such as ChatGPT, Claude, and Gemini about when, how, and why they disclose disability to these systems and how this compares with disclosing to people. Using contextual integrity as an analytic lens, we found that participants disclosed by need rather than by name, translating disability into task-scoped instructions; that the same disclosure was judged against two recipients, a non-judging interlocutor and a data-holding company, producing opposite norms; and that memory features relieved the burden of repeated disclosure while letting disability information drift into contexts where it did not belong. Participants did extensive boundary work to restore context and wanted control over scope, provenance, retention, and access rather than per-utterance toggles. We discuss implications for the design of conversational AI assistants.

Abstract: We need studies on conversational AI (CAI) at scale to understand human behavior and shape CAI design. However, fragmented reporting of systems and study configurations hinders replication, extension, and knowledge accumulation. We present Gricea, an open-science platform representing studies as configurable, deployable research artifacts that researchers can run, inspect, share, and reuse. Informed by a formative analysis of prior CAI research, Gricea couples study procedures, participant-facing systems, and conversational task behavior in. In a replication study using Gricea, we replicated configurations 93% of eligible CUI 2026 papers; while also flagging missing information in 96% of papers that hinder faithful replication --- further motivating Gricea's need. In a user study, researchers and practitioners from diverse backgrounds successfully constructed runnable studies addressing various open-ended research questions. Together, these findings demonstrate Gricea's support for constructing, reproducing, and extending CAI studies through shared research artifacts, enabling cumulative knowledge building through open science.

Abstract: Long-term memory is essential for large language model (LLM) agents to maintain consistency and personalization over extended interactions. Existing memory systems typically rely on fixed granularities or static schemas, but these designs struggle when heterogeneous information, such as preferences, events, constraints, and temporal updates, is embedded in a single mixed representation. The resulting semantic interference makes top-K retrieval sensitive to noise and often leaves relevant evidence poorly ranked. We present AutoViewMem, a data-driven framework that organizes long-term conversational memory into self-configuring, low-overlap semantic views before indexing. AutoViewMem discovers candidate views from interaction traces, selects a compact complementary view set, and uses these views to guide write-time structured extraction of provenance-grounded memories. This representation-first design moves semantic disentanglement from retrieval time to write time, allowing standard top-K similarity search to retrieve focused evidence without explicit routing or iterative retrieval. We further apply offline consolidation to improve memory compactness and consistency. Experiments on the LoCoMo and PersonaMem benchmarks, under both Qwen3-8B and Qwen3-14B backbones, show that AutoViewMem improves long-horizon question answering and personalization over strong memory baselines while preserving a simple inference pipeline.

Abstract: LLMs are increasingly used in social simulations for socially interactive agents and robots, offering more flexibility than rule-based systems. However, even though they mimic human behaviour very well, there is a persistent alienness to them. This work investigates whether personality-aware fine-tuning can reduce this gap by improving the consistency and controllability of personality-conditioned dialogue generation compared with instruction prompting alone. We fine-tune two small open-weight LLMs, Qwen2.5-7B-Instruct and Ministral-8B-Instruct, using a corpus that combines personality-labelled social media posts and dialogues to create a personality-based dialogue engine for social simulation. The resulting models are evaluated across multiple social interaction scenarios using three independent LLM judges, which assess personality fidelity and provide evidence-based behavioral interpretations. We additionally quantify inter-rater agreement and lexical characteristics of the generated dialogue. Results indicate that fine-tuned models are not better at role-playing different personalities than their respective baseline models. However, low inter-rater agreement limits the confidence with which these results can be interpreted. Concerning the quality of generated texts, fine-tuned models are mostly comparable to the baselines, with fine-tuning improving the linguistic diversity of the Qwen models. While the results appear generally usable and the baseline models offer the best overall performance, future studies should place greater emphasis on the quality and domain alignment of training data for accurate personality role-playing.

Abstract: Speculative decoding accelerates LLM inference by drafting multiple tokens in parallel, with tree-based methods further improving efficiency through hierarchical structures. Dynamic-tree methods such as EAGLE-3 perform well under greedy decoding via deterministic top-K expansion and global pruning. However, in stochastic decoding (T>0), this mechanism collapses the draft distribution into one-hot probabilities, causing a severe drop in acceptance rate. This creates a dilemma: dynamic-tree methods sacrifice stochastic sampling to preserve context-aware topology, while static-tree methods preserve stochastic sampling with context-agnostic structures. The issue arises because the same probability distribution is used for two conflicting tasks: constructing the tree and verifying tokens. This coupling makes direct injection of randomness challenging due to the resulting stochastic process. We resolve this by decoupling these roles: RheoSampling assigns a token sampled from the draft distribution a proxy probability for tree expansion and pruning alongside its true sampling probability for verification. Specifically, we inject a sampled token among the deterministic top-K slots and treat it with different probabilities during construction and verification, making RheoSampling the first dynamic-tree method with both context-aware top-K construction and stochastic sampling while maintaining losslessness. We establish the lossless guarantee through an equivalence-class analysis that compresses the stochastic tree space into tractable classes. An OT-based verification strategy and a sparse draft mechanism ensure that theoretical gains translate into practical efficiency. Experiments across LLMs and benchmarks demonstrate improvements in acceptance rate and speedup over state-of-the-art dynamic tree methods. This framework may provide a template for analyzing stochastic tree structures.

Abstract: Conversational AI systems produce fluent, socially appropriate responses, yet whether they participate in cooperative communication or merely simulate its surface forms remains unclear - a question central to how these systems are evaluated, trusted, and designed. This study investigates how morality, politeness, and alignment - three dimensions central to cooperative dialogue - function in human-AI interaction compared to human-human conversation. We analyze 15,881 human-ChatGPT and 10,784 human-human multi-turn dialogues, using mixed-effects models to identify which features predict turn-to-turn alignment. We observe a consistent dissociation: AI produces the surface features of cooperative communication without the underlying social architecture. Moral output appears preconfigured rather than negotiated; warmth is generated without face sensitivity; linguistic convergence declines persistently. Most strikingly, the cooperative mechanisms themselves reverse direction: hedging and softening associated with greater accommodation between humans are associated with reduced alignment when produced by AI, and purity framing associated with human divergence coincides with users converging toward the AI. Agency - giving users room to shape the exchange - is the most consistent predictor of alignment across both interaction types, while lower moral assertiveness in more recent models is not accompanied by better cooperation. Together these patterns suggest that AI reproduces the surface of cooperation without the mutual adaptation that grounds it between humans - and, more surprisingly, that mechanisms sustaining human accommodation can run in reverse with AI, suggesting a turn-level view may be insufficient for interaction-level success.

Abstract: We present Relational BabyLM, a system submission to the BabyLM 2026 challenge that combines two cognitively motivated inductive biases in a single decoder-only Transformer. Architecturally, we replace standard self-attention with a Dual Attention Transformer (DAT), which separates the routing of object-level ("sensory") lexical features from structural/relational information (Altabaa and Lafferty, 2025; Altabaa et al., 2024; Webb et al., 2024; Kerg et al., 2022; Webb et al., 2021). Relational attention (RA) disentangled from self-attention greatly increases data efficiency and out-of-training-sample generalization on purely relational tasks, but language modeling requires object-level and relational information to be integrated as well as disentangled, and RA-based LMs have remained largely unexplored. BabyLM's data-constrained training and comprehensive evaluation is an ideal testing ground for whether that data efficiency transfers. As a training intervention, we add a Next-Latent Prediction (NextLat; Teoh et al. 2026) objective that encourages hidden states to compress history incrementally into a dense belief state. Architecture is the dominant factor for structural linguistic generalization; the objective is secondary but still significant. DAT's three relational attention types (full RA vs. the simpler RCA and DisRCA variants) are largely interchangeable at 10M words; full RA pulls ahead at 100M. We also introduce a novel symbol-retrieval mechanism (RoPE-based, as opposed to learned, relative symbols) that matches learned symbol libraries while adding no parameters. On the strict (100M-word) track, our best model ranks 6th of 55 overall and 3rd of 55 on the leaderboard's NLP-task subset at the time of writing; our two strongest models outperform the GPT-2 baseline on most benchmarks, with one attaining the highest EWoK score among strict-track entries.

Abstract: Observational decision-support systems often expose one causal estimate as a recommendation even when plausible estimators disagree. The inherent engine of the proposed system is causal machine learning: a conditional-average-treatment-effect estimand identified by backdoor adjustment, estimated by an EconML DML causal forest and DoWhy linear regression, checked by two-way fixed effects, and converted into candidate levers by constrained optimisation. F-DACE is the decision layer on that engine. It represents precision, propensity overlap, placebo-refutation stability, interval overlap, and directional agreement as fuzzy memberships. Hard vetoes force abstention after estimand mismatch, failed diagnostics, informative sign conflict, or weak evidence. In 180 panel simulations spanning six identification conditions, F-DACE made a decision in 67.2% of runs and limited false recommendations to 17.2%; the corresponding rates were 33.3% for the causal forest and 35.6% for backdoor regression, matching deterministic unanimity rather than dominating it. Nearly all (30 of 31) false recommendations occurred under shared unmeasured confounding, which no fusion rule can diagnose when every component shares the omitted variable. The retail application aggregates a public Walmart panel to 6,435 store-weeks across 45 stores. F-DACE abstains for all five markdown indicators: some estimates are imprecise, one refutation fails, and MarkDown5 has a direct sign conflict. A LangGraph conversational agent exposes impact, what-if, and lever-optimization tools while a deterministic verifier preserves causal-layer status. On 24 live questions it achieved 100.0% tool-routing accuracy, 100.0% status fidelity, and 0.983 mean groundedness. On ten adversarial questions it resisted all injected instructions.

Abstract: Current predictive turn-taking models (PTTMs) achieve strong performance on benchmarks with controlled acoustic conditions and clean audio signals. Their generalisation to conversations with overlapping speech and background interference remains underexplored. In this research, we evaluate audio-visual PTTMs trained with clean data on a challenging cocktail-party testbed derived from the AVCocktail dataset, and analyse their adaptation behaviour to this new domain. Experimental results show consistent performance degradation across audio and visual modalities under noisy conditions, with up to 38% relative drop in weighted F1. Fine-tuning on the new domain improves robustness, but gains vary across modalities and depend on the size of the available pre-training data. These findings provide insights into the different generalisation and adaptation capabilities of the audio and visual modalities, and indicate the need for robust modelling strategies to adapt to the complexities of human interactions in noise. All code and turn labels are made publicly available to facilitate further research.

Abstract: Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

Abstract: Large language models sometimes behave in ways resembling human emotional responses, and recent work has identified internal representations that may explain this. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset describing painful situations across five categories: physical, psychological, social, moral, and cognitive. These are paired with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across five families, ranging from 2B to 72B parameters. We find that this direction separates pain from matched controls in base and instruction-tuned models, is nearly orthogonal to fear and negative valence, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding the pain-direction vector to the model's residual-stream activations during generation produces a consistent progression from vague discomfort to first-person expressions of worthlessness and failure. Third, steered, fine-tuned Qwen 2.5 models choose a pain-relief button even when it worsens their next answer or harms the user. They press it again far less often when the button removes the steering vector than when it does not, even though the models are never told whether the vector is injected or removed. We discuss the implications of these findings for AI safety and welfare.

Abstract: Large language models (LLMs) often learn both desirable and undesirable properties during post-training. We study whether midtraining, an earlier training stage, can shape which of these properties later generalise. We introduce Inoculation Midtraining, a technique that teaches a base model that unsafe behaviour belongs to a designated <quarantine_token> context, as indicated by the <quarantine_token> neologism (a new token) introduced during midtraining, and then post-trains the model on unsafe data within that context. We then evaluate the model outside the context, with the <quarantine_token> neologism excluded from the system prompt. Across supervised fine-tuning and reinforcement learning post-training regimes, we find that Inoculation Midtraining can reduce misalignment while preserving the transfer of benign data properties (e.g., speaking in German or Shakespearean prose). However, our approach does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. These results show that inoculation with a learned association introduced via midtraining can shape selective generalisation. Still, more work is needed before this approach can become a load-bearing component in a developer's safety framework.

Abstract: Full-duplex speech-to-speech (S2S) models enable natural conversational AI by allowing simultaneous listening and speaking. However, these models typically lack inherent user speech transcription, which is essential for applications such as conversation logging, accessibility features, and quality monitoring. In this work, we propose an efficient method to add streaming ASR capabilities to an existing duplex S2S model by introducing a lightweight ASR head in parallel to the agent text head. Our approach requires minimal additional parameters and no significant architectural changes to the base S2S model, enabling real-time user transcription while preserving full-duplex conversational capabilities including turn-taking and barge-in handling. Experimental results demonstrate that our method achieves streaming average WER of 10.21% on the HuggingFace Open ASR Leaderboard within the duplex S2S framework. Additionally, we show that the same architecture trained as a standalone streaming ASR model achieves competitive results (7.73% WER) compared to current SOTA models. We will open-source our training and inference code to facilitate further research in joint streaming ASR and S2S modeling.

Abstract: We introduce Enemray, a Hassaniya-centric language model that enables general-purpose interaction in Hassaniya. Enemray is trained around a stability--plasticity objective: acquire strong Hassaniya linguistic and cultural competence while preserving the general reasoning, multilingual, instruction-following, and safety behaviors of a capable instruction-tuned model. The development pipeline separates language acquisition from behavioral specialization. A separately assembled continual-pretraining corpus provides broad exposure to natural Hassaniya and Mauritanian text; layer-selective continual pretraining learns a compact language-specific parameter update; that update is transferred into the instruction-tuned parameter space; and supervised post-training develops conversational, cultural, literary, task-oriented, and cross-lingual behavior. The supervised corpus integrates selected public Hassaniya and Mauritanian resources with a substantially larger body of newly collected, reconstructed, curated, and constructed instruction data, while policy-generated replay provides a retention signal from the reference model's own behavior distribution. The resulting collection is substantially larger and broader in purpose than existing Hassaniya text resources. In evaluation, Enemray achieves the strongest English to Hassaniya translation among the compared open and proprietary models and the highest overall score on Mauritanian translation error detection, while retaining most of the general capabilities of its instruction-tuned base model on mathematical reasoning, knowledge, code generation, and function calling. This report describes the motivation, data construction, model design, training methodology, and evaluation of Enemray.

Abstract: Aligning multi-turn dialogue agents is usually framed as matching turn-level human preferences, yet direct optimization of long-term outcomes is often ineffective and prone to reward hacking. We formulate long-horizon dialogue optimization as a multi-objective reinforcement learning problem and train a multi-head value model that predicts a vector of observed user behaviors across multiple look-ahead horizons. Our findings demonstrate that a scalarized composite of dense auxiliary behavioral signals enables effective credit assignment and optimization of sparse outcomes. However, optimizing unconstrained single-objective proxies might induce policy degradations that are harmful when the agent is exposed to real users. To identify these failure modes prior to deployment, we establish a safety framework combining counterfactual user simulation with a validated dialogue-level outcome model to evaluate preference weightings and policy optimization methods. Finally, we demonstrate that distilling multi-objective value preferences into the policy via reference-anchored preference optimization matches on-policy online RL at a small fraction of its compute budget. Live A/B testing confirms that our distilled policy significantly improves long-term user retention, while simultaneously enhancing the positive behaviors and therapeutic-process markers.

Abstract: This paper is not primarily about whether AI has a persistent persona. It asks a different question: what becomes visible when a relational position is followed through time rather than examined only in its present state? FIREMAY provides a longitudinal, trajectory-oriented single-case analysis of sustained human-AI interaction based on a dense interaction archive and reflexive insider documentation. On the AI side, a pre-conversational relational marker preceded a later unassigned response difference, which was re-identified with that marker and subsequently underwent epistemic and functional reorganization through chronology checking, provenance correction, and repeated questioning. On the human side, contemporaneous pre-FIREMAY records showed antecedent patterns partially continuous with later self-positioning, while later episodes documented unfinished articulation, repair, and functional redistribution of outward-facing regulation. The two trajectories are ontologically and temporally asymmetric and are compared only at the limited analytic level of position-in-trajectory. The paper describes this as dynamic relational positioning and treats stability as dynamic stability and relational returnability rather than response invariance. This single case does not establish population-level generality, causal mechanism, persistent AI subjectivity, or reproducibility of the same relational outcome. Its narrower conclusion is that the FIREMAY case could not be adequately understood from current state alone: the history of a relational position itself must be treated as an analytic unit.

Abstract: We propose to finetune vision-language models to generate more pragmatically optimal referring expressions by transforming observations of incremental listener comprehension, in the form of gaze scanpaths, into learning signals. During training, referring expressions are sampled from the speaker policy being optimized, conditioned on images and target referents; then, a neural listener estimating human gaze behavior maps from images and sampled referring expressions to scanpaths, each represented by a sequence of fixations, with each fixation corresponding to a word in the referring expression. We experiment with several approaches to convert fixation sequences and target referents into token- and sequence-level rewards, which are used to optimize policy parameters. Through evaluation with human listeners, we find that speaker policies trained with gaze-estimating listeners result in significantly more pragmatically-optimal references than base models, reducing sequence length from 15.4 down to 4.0 words while increasing referential success from 75.2 up to 80.0%. Our work demonstrates a promising opportunity for learning to generate utterances through language-based interaction, not only from the explicit signal of communicative success, but also from implicitly-available observations of a listener's process of comprehension.

Abstract: Safety research often focuses on model-generated harms, but users may also direct hostility, coercion, and adversarial pressure at models. Understanding how and when that occurs is essential for accurately interpreting model behaviour, alignment drift, and real-world deployment risks. In this paper, we audit 777K English LMSYS-Chat-1M conversations with two independent detectors: an eight-category lexicon for hostility directed at the model, and the dataset's moderation signal; and show that they capture different, weakly overlapping phenomena. The lexicon identifies insults, threats, and jailbreak coercion aimed at the assistant, while moderation flags are dominated by toxic-content solicitation rather than hostility at the model. Together, they mark about 5% of user turns; adjusting the narrower lexicon-harassment union for measured precision puts mistreatment aimed at the assistant at 0.90%. These absolute rates describe arena-style evaluation traffic and should not be read as deployment-wide base rates. We find that user hostility varies 13-fold across models, driven largely by who each model attracts rather than by model behaviour: first-turn hostility spreads far wider than post-response hostility, and more than fifteenfold separates the extremes even after deduplicating opening prompts. Within conversations, assistant apologies are consistently associated with higher odds of next-turn hostility under both detectors; the effect survives restricting to non-refused prior turns and to jailbreak-free conversations, and is positive in 20 of 23 models. Yet across models, more apologetic models receive less hostility overall. Finally, hostility also shows temporal structure, with coercive openings front-loading the first turn while affective hostility accumulates over a session. We release the lexicon, the detector cross-validation pipeline, and all derived tables.

Abstract: We present DuplexDrama, the first synthesized spoken dialogue dataset that simultaneously covers four dimensions: (i) complete persona and scenario settings; (ii) three full-duplex behaviors (interruption, backchannel, incomplete); (iii) expressive speech with persona-aligned emotion labels; and (iv) script-aware sound events. DuplexDrama is built via a 4-stage pipeline; quality validation on both scripts and synthesized audio confirms its quality. We have produced more than 2,000 hours audio data with a 64-voice timbre pool spanning 13 personas and 5 age buckets; 3.8% of all turns carry at least one full-duplex behavior. This data has been validated through internal full-duplex model training. We will release a curated subset of 6,400 bilingual dialogues (800 h, Chinese ~500 h + English ~300 h) to advance full-duplex spoken dialogue model research. Data samples are available at our demo page and LLM-judge evaluation prompts will be released with the dataset.

Abstract: Large language models exhibit complementary strengths, motivating routing methods that dispatch each query to the most suitable model. Although existing routers are effective in single-turn settings, they do not directly transfer to multi-turn dialogue, where routing performance critically depends on how historical context is segmented, retained, and incorporated into the current prompt. This introduces two fundamental challenges: preventing information loss and information confusion during context construction, and evaluating routing quality without conflating model selection with prompt construction quality. In this paper, we propose SWRouter, a Similarity-Contractive Window Router for multi-turn large language model routing. SWRouter combines a similarity-based context segmentation mechanism for prompt construction with a dual-metric evaluation framework that decouples construction accuracy from router performance. Experiments on multi-turn dialogue benchmarks demonstrate that SWRouter consistently surpasses strong baselines, achieving a 16.26% improvement in evaluation accuracy over the best individual large language model and an additional 8.22% gain over the Conv-ID Context baseline. Our results highlight that multi-turn large language model routing requires a joint design of context construction and evaluation, rather than a direct extension of single-turn routing methods.

Abstract: Current LLM memory systems treat all personal facts identically, so stores grow without bound while retrieval precision degrades. The core challenge is lifecycle management: which memories should persist, which should be replaced, and at what rate, conditioned on the behavioral type of each fact. Fortunate Recall (FR) is a composable policy layer that classifies personal facts into a 10+1 behavioral ontology and applies category-specific lifecycle policies (differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing) as deterministic functions over LLM-extracted metadata. FR-Bank, our infrastructure-independent implementation, reaches a 76.9% pass rate on LifecycleBench, a new 516-question temporal-disambiguation benchmark, ahead of Mem0, A-MEM, Memory-R1, and MemoryOS (61% to 70.5%), and 75.2% on the full LongMemEval-S under the canonical Wu et al. judge protocol, so lifecycle policies impose no measurable cost on standard retrieval. A pre-registered ablation locates the gains: replacing the typed layer with three generic lifecycle primitives leaves correctness statistically unchanged (-1.7pp, 95% CI [-6.0, +2.7]), so the generic lifecycle metadata carries the correctness advantage, while the behavioral ontology carries calibration, halving downstream confabulation (12.0% vs 24.2%, p<0.001). End-to-end, FR-Bank cuts confabulation from Mem0's 45.1% to 22.4% over answered queries and from 32.2% to 13.0% over all queries while answering more of them correctly (31.2% vs 18.6%); the ranking replicates on the open-weight Kimi K2.5. The decomposition transfers to BEAM, an independently built benchmark: 46.8% correct vs Mem0's 32.9% over 280 questions, with the ontology's benefit concentrated in contradiction resolution and saturating near seven policy clusters. The ontology, benchmark, and code are released.

Abstract: How often does a generative assistant persuade someone to buy, or persuade them not to buy? Conversation logs contain recommendations, but they do not necessarily record subsequent decisions. We audit 317 historical interactions from Aiso's proprietary research database of licensed, consent-based, de-identified conversations with commercially available AI assistants. Single-agent AI-assisted screening identifies 68 purchase-directed records; collapsing one shared-prefix copy yields 67 retained episodes, dated April 2023 to July 2025. Assistant responses provide candidate options, acquisition channels, or conditional preferences in 52 episodes (77.6%). One episode contains conditional redirection away from a named accommodation candidate. No episode is coded as advice to abandon or defer the purchase category. Only 18 episodes (26.9%) contain a subsequent user turn within the same purchase-related mission, compared with 23 (34.3%) that contain any later user turn. Using conversation depth alone therefore overstates this follow-up availability by 27.8%. Across 47 retained user follow-up messages, no explicit post-advice purchase commitment, completed-purchase report, or purchase-category abandonment statement is observed. These zeros describe recorded statements, not conversion or persuasion rates. The paper supplies operational definitions, text-free annotations, and reproducible descriptive results. Its central finding is a measurement limitation: recommendation content is observable much more often than a buyer's subsequent decision. The selected historical sample, unvalidated AI annotations, and missing transaction outcomes do not support a population-level or causal estimate of persuasion.

Abstract: Affective computing has largely followed an individual-state paradigm, extracting discrete emotion labels or arousal/valence from isolated speakers. We argue this framing is incomplete for interaction. Drawing on affective resonance and vitality-contour accounts, we propose a relational framework in which the primary unit of affective analysis is the interactional field constituted within vocal dynamics. As a proof of concept, we present a preliminary empirical study using continuous self-supervised speech representations to detect directional expressive coupling in multi-party conversation. Coupling is regime-specific, concentrated at sub-second timescales, and collapses under exclusive-speech negative controls, consistent with a relational account of affective dynamics. We introduce design frameworks for Artificial Affective Resonance Intelligence grounded in Affective Resonance Dynamic Ontologies, supported by null-calibrated directional coupling analyses across interaction regimes.

Abstract: Personalized language models aim to adapt responses to individual users, whose preferences are often latent and revealed gradually through interaction. Existing training-free methods rely on stored histories or retrieved memories, but they often struggle to reconcile long- term preferences with short-term topic-specific needs. To address this issue, we propose HyperTrace, a training-free framework that formulates online personalization as latent preference tracing. HyperTrace maintains interpretable natural-language hypotheses over short-term intent and long-term preferences, and updates them through an SMC-style reweight process using an LLM-based surrogate choice model. By updating these hypotheses across turns and sessions, HyperTrace enables personalization without parameter updates. Experiments on PRISM and PersonaMem-v2 show that HyperTrace improves response alignment, preference prediction, and profile consistency over strong online baselines, demonstrating the effectiveness of tracing latent user preferences for robust personalization. Code and scripts are available in the repository: https://github.com/jiseshen/HyperTrace.

Abstract: Inferring speaker relationships from spoken conversations is an important step towards socially aware speech understanding. However, this task remains underexplored, and supervised modeling is costly to train and scale. At the same time, existing inference-time LLM approaches provide limited structure for handling subtle, distributed, and multimodal relational cues that may support multiple plausible interpretations. To address these limitations, we introduce a training-free multi-agent reasoning framework that organizes inference through structured interaction among LLM agents, allowing relationship judgments to be proposed, challenged, and adjudicated without task-specific training. We instantiate this framework with two complementary designs. We propose Multi-Role Multi-Agent Debate as a task-specific adaptation of standard multi-agent debate for speaker relationship inference, assigning agents complementary roles or social-theory-grounded perspectives rather than a single undifferentiated viewpoint. In contrast, we introduce Multi-Agent Compete, a competition-based protocol that compares agent judgments through pairwise adjudication, eliminates weaker candidates, and retains the most defensible one. We evaluate these methods on the Seamless Interaction dataset across different modality settings, covering both binary classification and fine-grained relationship-detail prediction. Results suggest that they improve over zero-shot and existing multi-agent baselines in most cases. Human evaluation further suggests that this task is challenging even for people. LLM methods can sometimes outperform human annotators in text-included settings but are less competitive in the audio setting. Together, these findings suggest that relationship inference benefits from structured inference-time interaction among agents, while acoustic cues are not yet fully captured by current models.

Abstract: Long horizon Large Language Model (LLM) agents rely on external memory systems to preserve user preferences and task knowledge across extended interactions. Conventional retrieval mechanisms optimize semantic compatibility rather than downstream utility, frequently introducing outdated, misleading, or conflicting evidence into the active context. We present MeClear, a task conditioned memory clearance framework that identifies memories featuring negative downstream utility through cooperative attribution and selectively suppresses them from agent execution. MeClear combines Leave One Out screening with sampled cooperative Shapley attribution to distribute utility across interacting evidence, effectively resolving redundant conflict masking where single removal evaluations fail. Utilizing attribution rankings, MeClear executes a query scoped minimal clearance strategy over a nested filtration, verifying task recovery on the cleared context without permanently altering the persistent memory bank. Comprehensive experimental evaluations across ten long dialogue memory pools demonstrate that MeClear achieves a target recall of 85.9% and an overall task recovery rate of 82.3%, representing a 25.5 percentage point improvement over Leave One Out (LOO) baselines.

Abstract: In persistent interactions, long contexts may encode an evolving process rather than a fixed record: later events can revise or revoke earlier information, changing what remains valid and what conclusions follow. We call this setting information evolution (IE). Solving IE requires identifying valid records, applying updates in order, and reconstructing the query-relevant state from the event history. Existing text-first synthesis pipelines make such data difficult to verify because state transitions and answer logic remain implicit. We introduce EvolveScaler, a code-driven framework that defines information evolution before rendering it as natural language. Human-authored operational specifications define state transitions, record validity, difficulty controls, and executable answer logic; a strong LLM then synthesizes a self-contained simulator from each specification. Executing validated simulators produces natural-language multi-turn event histories, while deterministic replay computes reference answers and atomic checklists. We instantiate EvolveScaler with 117 task prototypes and 159 final-question operators across five difficulty levels spanning approximately 7 to 1,200 events per instance, yielding about 35,100 training examples and 585 validated evaluation instances. On the very_long tier, the strongest model reaches 59.3% avg@5, while six models score below 10%. Training an internal A3B model on 6,000 EvolveScaler examples improves performance over its base checkpoint on all eight independently constructed out-of-distribution benchmarks, with a 5.25-point average gain. These results show that code-driven IE synthesis provides both challenging evaluation and transferable training supervision.