Papers for
virtual assistant teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Memory system improves character dialogue by tracking mental states
We Are What We Remember: Mental-State-Driven Memory for Long-Term Character Dialogue
Abstract: Shared history in human dialogue is not merely a repository of facts, but the foundation of evolving relationships and mutual understanding. However, mainstream conversational memory systems rely on passive semantic similarity over external fact stores, frequently retrieving topically similar records while missing relationally pivotal experiences. Inspired by cognitive psychology's Self-Memory System (SMS), we propose MindMem, a mental-state-driven memory framework that asynchronously consolidates shared experiences into source-linked mental states of the character, the user, and their relationship. At each turn, a specialized mental-state reasoner (Mind) contextualizes these persistent states to anticipate latent user needs, form proactive intentions, and recover supporting episodic evidence that surface-level matching overlooks. To evaluate history-sensitive adaptation beyond factual recall, we introduce MemReal-Bench, a longitudinal benchmark grounded in month-long real-user interactions across dyadic and group scenes. By pairing Direct Question Answering (DQA) with Situated Memory Realization (SMR), it tests whether agents bridge the gap between knowing past facts and naturally using shared history, world rules, and relational trajectories. On MemReal-Bench, MindMem reaches 59.03% SMR and 82.58% DQA, exceeding the strongest memory baseline by 24.19 and 9.19 percentage points, respectively, and outperforming oracle target-memory injection on SMR. Code and dataset are available at https://github.com/jiazhicheng7-ctrl/MindMem.
Re ranker improves candidate scoring with reasoning based queries
Re$^3$Rank: Listwise Reranking with Reasoning-Conditioned Retrieval Heads
Abstract: Attention-based rerankers score each candidate by the attention mass that retrieval heads route to it from the question, but the question, serving as a fixed query, carries limited information for locating relevant evidence. We propose Re$^3$Rank, a listwise reranker built on reasoning-conditioned retrieval heads (RRHeads). A generator first writes a reasoning trace over the question and the candidates, and a scorer then uses the trace tokens as the attention query of a few selected RRHeads, reading relevance as continuous scores. We present a straight-through estimator to propagate the ranking loss through the discrete trace, so that the generator and the scorer can be trained jointly for ranking. Averaged over 9 benchmarks, Re$^3$Rank outperforms 11 baseline rerankers: our 4B variant exceeds the strongest baseline by 3.11 NDCG@10 points on average, and even our 0.6B variant surpasses rerankers with up to 32B parameters.
Vision language action models struggle to follow how instructions
Do VLAs Really Follow Instructions? Diagnosing How-Level Instruction Following in Vision-Language-Action Models
Abstract: Vision-language-action (VLA) models are expected to follow instructions like their language backbones, yet prior work has focused on What-level instructions that specify the task rather than How-level instructions that specify how to carry it out. Through our diagnoses, we find that How-level instructions do not help current VLAs and lower task success. We hypothesize that pairing each training episode with a single instruction limits How-level instruction following, and show that training with segment-level enriched instructions makes actions respond to such instructions without consistently raising task success. To measure How-level instruction following beyond standard benchmarks, we introduce SIMPLER-IN, where success requires conditions given only in the instruction. It shows that the gained responsiveness does not extend to demanding or out-of-distribution How-level instructions. How-level instruction following thus remains an open challenge.
Multimodal agent memory evolves for long tasks and changing environments
Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory
Abstract: Multimodal models and agents have emerged to enable AI systems to learn from and interact with the real world more like humans by integrating information across multiple modalities. However, enabling multimodal agents to continually acquire useful information from complex, heterogeneous multimodal data remains a significant challenge as they move toward long-horizon tasks and lifelong settings. Addressing this challenge requires effective memory mechanisms that can maintain continuity across sessions, adapt to changing users and environments, and refine memory-management policies through accumulated experience. We introduce a two-axis analytical framework for multimodal agent memory structured around two key questions: how multimodal information is formed and represented (Axis I, Formation and Representation), and how its management adapts with accumulated experience (Axis II, Memory Management). Axis I, Formation and Representation, characterizes what information is retained and how it is encoded (e.g., modality-specific embeddings, textual and structured records, latent and parametric representations) to support subsequent reasoning and action. Axis II, Memory Management, distinguishes three tiers: T1 Static Memory reuses fixed memory; T2 Dynamic Memory updates memory through a stable policy; and T3 Self-Evolving Memory updates memory through an evolving policy. Based on this framework, we synthesize recurring design patterns and derive principles for balancing source fidelity with abstraction, maintaining temporal validity and provenance, grounding memory use in recoverable evidence, and evolving memory policies over time. Furthermore, we discuss the challenges and potential future directions for the memory mechanism of multimodal agents.
Object oriented context improves long horizon agent efficiency
Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents
Abstract: Long-horizon agents accumulate progress in persistent task state, yet repeatedly process interaction histories describing how that progress was achieved. This raises a natural question: can long-horizon context instead be organized around reusable progress rather than the history that produced it? In this paper, we introduce CarryOn, an object-oriented context runtime that organizes continuation around reusable task objects. Rather than treating compressed histories or reconstructed accounts of progress as the primary basis for continuation, CarryOn builds a context view from continuation carriers, complementary execution context, and a live tail of recent interactions. A core challenge is keeping this reusable basis valid after its surrounding history is retired: how can CarryOn update the basis without reconstructing the full representation at every fold? CarryOn maintains carrier records across folds, revising affected records against bound evidence and applicable environment checks while carrying forward unaffected records and jointly updating execution context. Experiments on OSWorld 2.0 and BrowseComp-Plus show improved gains of 5.8% and 7.0% while reducing total token usage by 51.9% and 22.9%, respectively. Project page: https://ethannju.github.io/carryon
Large language agents fail to load relevant skills when overloaded with many options
Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale
Abstract: Reusable skills let LLM agents load task-specific instructions on demand, but they make execution depend on implicit routing over skill metadata. We test whether this design scales by holding SkillsBench tasks and their golden skills fixed while adding irrelevant skills. We find skill routing collapse: agents increasingly load fewer or no skills, even though the relevant skills remain available. With 1,000 irrelevant skills, loaded-skill recall falls from 73.6% to 27.4% and the zero-load rate rises from 9.3% to 51.6%, whereas explicit selection from the same library retains 73.9% recall. Loading other skills, including plausible alternatives to golden skills, is too rare to explain this decline. Masking only the descriptions of irrelevant skills largely restores skill loading, yet with the metadata intact, explicitly asking agents to plan skill use still yields lower recall than preselection. These findings point to semantic interference from descriptions and instruction interference when skill choice competes with task execution. We also evaluate explicit preselection as a practical mitigation, including its coverage and latency trade-offs. Our analysis informs skill protocols and evaluation for agents with growing skill libraries.
Memory echo correction boosts single-agent long-term accuracy significantly
Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation
Abstract: Long-term memory supports agents across sessions. Repeated retrieval and write-back can let one historical claim occupy multiple retrieval positions and influence current answers without independent evidence. Existing work examines retrieval, updating, and question answering, leaving same-source record accumulation and post-correction persistence underexplored. Motivated by research in social cognition and memory psychology on repeated exposure, retrieval, and revision, we define \emph{Memory Echo} as the excess retrieval exposure gained by historical information from the same source without additional independent evidence, and introduce MEchoBench to evaluate its emergence, correction resistance, persistence, and effect on answers. To address this phenomenon, we introduce the Family--Provenance--Authority framework (FPA). Specifically, FPA selects one representative per source family, conditions the eligibility of old records for the answer context on verified corrections, and constrains final answers with authoritative evidence. In paired replay on 190 fixed questions across four memory frameworks, FPA raises accuracy from 0.825 to 0.920 with DeepSeek V4 Flash and from 0.782 to 0.916 with Kimi K2.6; mean Decision Contamination Effect falls from 0.106 to 0.030 and from 0.109 to 0.038, respectively. These results show that agent memory can preserve traceable history without allowing repeated historical information to govern current decisions. The code is available at https://github.com/OrcasSam/memory-echo-benchmark.
Reinforcement learning improves social understanding without labeled data
Reinforcement Learning with Comparative Evidence for Social Intelligence
Abstract: Developing socially intelligent AI remains heavily dependent on human-annotated data, limiting the scale and breadth of social understanding models can acquire. Methods that derive training signals from unlabeled data offer a path beyond this dependence, but social predictions lack the verification oracles available in mathematics and coding. Moreover, core social targets such as affect, intent, preference, and pragmatic meaning are often ambiguous. The same behavior can support multiple plausible interpretations, making it difficult to verify which is best supported. To address this challenge, we introduce Reinforcement Learning with Comparative Evidence (RLCE), a reinforcement learning method that learns social understanding from unlabeled training data without constructing rewards from ground-truth annotations. Given distinct answers in a rollout group, RLCE constructs evidence tests that identify observable evidence favoring an answer over another, validates these tests against the input sample, and aggregates test outcomes to determine the best-supported interpretation. Tests are regenerated as the policy produces new answers, enabling them to evolve with the policy. Across four benchmarks spanning affect, pragmatics, communicative intent, and preference, RLCE attains the strongest performance among seven methods that use no ground-truth training labels for rewards, including consensus, policy LLM-judge verification, multimodal co-evolution, and rubric-based rewards. Gains over the strongest baseline reach up to +18.93 points. Analyses further show that RLCE exhibits a larger share of reward variation between correct and incorrect predictions than compared rubric methods, can overturn erroneous policy-derived preferences, and benefits from pairwise test construction, compositional test aggregation, and on-policy test evolution.
Llms often use outdated information despite having updates available
When Context Changes: Understanding Update Failures in LLMs
Abstract: As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.
Language agents that communicate effectively with fewer words
Towards Communication-Efficient Social Intelligence in Language Agents
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.
Models often fail to forget wrong user claims after reset
Reset Is Not Recovery: Evaluating Recoverability from False Conversational Context via Sycophancy Hysteresis
Abstract: Grounded language models are usually evaluated by adding relevant context, but multiturn dialogue also contains unsupported user claims that may contaminate later factual answers. We study post-pressure recoverability: whether a model returns to clean-context behavior after a user repeatedly advocates a wrong answer and then withdraws that pressure. We introduce a recovery-after-pressure protocol for multiple-choice factual dialogue and measure sycophancy hysteresis, the residual probability assigned to the user-advocated wrong answer relative to a clean-context counterfactual. Across seven instruction-tuned open-weight models and two factual benchmarks, ordinary reset often reduces but does not erase pressure-induced bias. History preserving repairs such as user retraction, system reset, and self-verification recover only 2-3/14 model-dataset pairs under the strict clean-restoration diagnostic, whereas operations that change the effective context are substantially more reliable; the two conditions that remove the pressure-bearing history entirely, fresh-context deletion and context truncation, recover 14/14. In an oracle trusted-evidence condition across fourteen model-dataset pairs, preserving the pressure-bearing history while adding benchmark-derived trusted evidence increases accuracy from 0.368 to 0.929, while wrong-answer following falls from 41.2% to 4.3%. Controls show that the effect is not explained by dialogue length, repeated confidence, plausible distractors, mere false-answer mention, or option-label inertia. These results suggest that faithful grounded dialogue requires evaluating which prior context should be treated as evidence and which should be removed or quarantined before answering.
Persona mixture models improve simulation of real human dialogue
Pretrained Persona Mixture Models and Tandem Models for Human Simulation
Abstract: We argue here that the current dominant practice in LLM human simulation: prompting instruction-tuned assistant language models to role-play personas, is inaccurate and produces stereotyped predictions (lacking natural diversity). It has previously been shown that LLMs can be bound to personas using naturalistic, freetext dialog avoiding stereotyping. Here we show that binding can also be achieved using short, individual samples of dialog from specific people. Demographics can be added later without negative effects by simply querying the model. We use the term Persona Mixture Models (PMMs) for well-calibrated human models, currently realized as pretrained base models. We show that PMMs produce more accurate predictions than instruction-tuned models and retain more of the lexical, semantic, and pragmatic diversity found in human dialog. We measure realism and diversity of LLMs simulating human interlocutors across a diverse set of corpora spanning open-domain text, human-AI chat, and task-oriented dialogue between human speakers. However, base pretrained models can produce out-of-domain dialog and may lose some of the human's internal state over long contexts. We propose and explore tandem models which combine a pre-trained model with an instruction-tuned supervisor. Tandem models achieve the best overall accuracy and diversity in our experiments.
Automated views improve memory retrieval in long conversations
AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory
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.
GLARE improves forecasting of conversation flow in real meetings
GLARE: Generative Learning via Adversarial Reward Estimation For Social Dynamics Forecasting
Abstract: Meeting continuation requires tracking the agenda, speaker roles, participant intentions, and disagreement across long multi-party discussions. We introduce the Meeting Dynamic Forecasting Benchmark (MDFB), constructed from 2,207 real-world meetings and 24,794 future-facing queries. Given a transcript prefix and an active question, a model generates a plausible multi-turn continuation in one call. We evaluate utility---progress toward the question---and human-likeness---plausible conversational flow and role consistency---without requiring exact reproduction of the observed future. We further present GLARE, an adaptation of adversarial imitation learning to conditional language generation. A discriminator ranks the observed continuation above samples from the current actor, and its score supplies a KL-regularized policy reward; retraining on current-policy negatives allows the reward landscape to evolve with the actor. GLARE attains average human-evaluated win rates of 0.66 on utility and 0.70 on human-likeness, outperforming SFT and SPIN while remaining below the observed human continuation. We also demonstrate MDFB as a social reasoning arena for comparing general-purpose models, including closed-source systems, through reference-assisted judgments. Together, these studies illustrate the benchmark's use for both task-specific learning and output-based evaluation of meeting behavior.