NeurIPS 2026 Papers

Large language models (LLMs) are increasingly deployed as interactive agents, yet their capacity for social and strategic reasoning over extended interaction remains poorly understood. Existing evaluations rely on static vignettes or single-game benchmarks that cannot capture the sustained, multi-faceted reasoning that real-world multi-agent settings demand. We introduce Mindgames, a multi-game arena and evaluation platform for LLM agents that operationalizes complementary reasoning demands relevant to ``theory of mind'': belief attribution under hidden information, opponent modeling through repeated strategic interaction, cooperative inference under knowledge asymmetries, and sustained deception in social deduction. Built on TextArena, Mindgames provides a unified interaction interface, TrueSkill-based rating, and full trajectory logging across four game environments. We instantiate Mindgames through a 2025 competition cycle hosted at a major AI conference, which assessed 944 submitted agents from 76 teams across four games: Colonel Blotto, Iterated Prisoner's Dilemma, Codenames, and Secret Mafia. Our analysis surfaces both agent-level and evaluation-level limitations: brittle rule adherence remains a major bottleneck, top-performing systems repeatedly rely on explicit structural scaffolding, and leaderboard validity differs sharply across environments. In particular, failure-heavy environments can reward robustness to opponent errors as much as strategic ability, with Secret Mafia exhibiting a pronounced error-survival confound in this cycle. We release a dataset of 29,571 multi-agent games (comprising 94,132 player trajectories and 243M tokens) with turn-level observations, actions, and rewards, together with MG-Ref, a deterministic offline tournament protocol that scores new agents against a frozen reference pool of top-ranked, low-error Stage~II submissions under the same error-attribution lens used in this analysis. By releasing the benchmark interface, dataset, MG-Ref, and starter kit, Mindgames provides a reproducible resource for studying progress in multi-agent social and strategic reasoning.

In cooperative multi-agent reinforcement learning under partial observability, agents typically differ in both their local observations and in how much value those observations let them recover. Most existing methods treat agents symmetrically when shaping credit and exploration, ignoring this asymmetric information disadvantage. We introduce the per-agent value gap, a state- and agent-specific quantity that measures the value loss attributable to local information limits, and show that it serves as a single, measurable signal that unifies credit assignment and exploration: down-weighting policy updates where an agent is at a strong disadvantage, and rewarding behaviors that reduce the disadvantage along the trajectory. We instantiate this idea as \method{}, which estimates the gap online with a per-agent local value head and uses it to drive both advantage shaping and an information-seeking intrinsic reward. We further show that the resulting shaping reduces the variance of the policy-gradient estimate at high-disadvantage states under a mild correlation assumption, and that the intrinsic bonus vanishes in expectation under any stationary optimal policy, so the augmented objective remains consistent with the original cooperative objective. On sixteen tasks across five cooperative benchmarks, \method{} ranks first or statistically tied for first on fifteen of sixteen tasks (strictly first on twelve), with the largest gains on settings combining sparse rewards and asymmetric observability (e.g., a $24\%$ mean relative gain over the strongest baseline averaged across the three SMAC v2 tasks).

Memory-augmented LLM agents tackle complex long-horizon tasks by recursively summarizing interaction trajectories into compact memory. However, existing approaches typically train these memory policies using outcome-based reinforcement learning, failing to localize where intermediate memory quality degrades. As interactions unfold, ambiguous recursive summaries progressively discard or blur task-relevant information. This exacerbates belief deviation, obscuring the agent’s estimate of the latent task state and ultimately derailing long-horizon reasoning. We therefore argue that memory optimization should focus not merely on trajectory-level success, but on the clarity of the belief induced by intermediate summaries. To this end, we introduce Belief Entropy, a self-supervised proxy that probes how uncertain the model remains about the latent task state given its current memory. Based on this proxy, we propose Metacognitive Memory Policy Optimization (MMPO). Instead of relying only on sparse outcome-based signals, MMPO provides fine-grained, memory-specific supervision by explicitly penalizing summaries that induce high epistemic uncertainty. Experiments show that MMPO consistently outperforms existing methods on diverse long-horizon tasks, maintaining 97.1% performance even when scaled to 1.75M-token contexts.

Self-improving LLM agents typically improve answers, not the $\textit{process}$ by which answers are improved: their search loops and mutation rules are held fixed by design. We introduce $meta^n$, a framework for $\textit{recursive}$ self-improvement in which a single universal meta-operation $\Omega$ is applied repeatedly to build a hierarchy of meta-layers whose depth is sized by convergence rather than prescribed. Each layer observes execution traces from below, diagnoses failure patterns, and emits executable code (strategic guidance via pre-processing and reusable functions via a code library) whose effects cascade downward. Inter-layer conditioning permits up to $\prod_{d=2}^{n} k_d$ effective strategy-tactic combinations across $n$ layers rather than $\sum_d k_d$. An evolutionary orchestration mode maintains a growing archive of candidate chains with fitness-weighted parent selection and cross-candidate inspiration. Across seven benchmarks (combinatorial optimization, two text classification tasks, terminal agent tasks, mathematical discovery, symbolic regression, algorithm speedup), $meta^n$ beats prior self-improving agents on all five (of seven) benchmarks where they apply, with the largest gap on CO-Bench (0.849 test vs. Gödel Agent's 0.000 and OpenEvolve's 0.816). Layers specialize without prescription and their roles emerge at depth: generic primitives at depth 2 propagate as cross-task callables, task-aware routing emerges at depth 3, and deeper layers refine via specialization. Code and data will be released publicly.

Large foundation models have accelerated progress toward general-purpose agents that interact with humans and other agents through language and multimodal signals. However, robust multi-agent decision-making requires reasoning about what other agents know, intend, and are likely to do under partial observability. Current agentic systems often operate through prompt design, memory, or end-to-end behavioral shaping, but typically do not learn an explicit partner-state representation that can be reused as a decision variable across tasks. We introduce mental-model-enabled agents, a framework that equips an agent with a latent mental model of its counterpart, allowing it to infer hidden beliefs, intentions, and likely reactions from the observed history and use these inferences to guide action selection. Our method learns an amortized recursive Theory-of-Mind representation, with first- and second-order mental-state structure, jointly with a belief-conditioned reward model that evaluates candidate actions relative to the inferred partner state. A policy is then learned under this belief-aware signal, yielding an agent that can act independently at inference time while retaining the benefits of explicit partner modeling. We evaluate the same framework on both language-only and multimodal benchmarks. Across these settings, explicit mental-state modeling consistently improves interaction quality and Theory-of-Mind performance over base agentic systems, showing that structured partner modeling is a useful inductive bias for general multi-agent systems.

Memory-augmented LLM agents enable interactions that extend beyond finite context windows by storing, updating, and reusing information across sessions. However, training such agents with reinforcement learning in multi-session environments is challenging because memory turns the agent's past actions into part of its future environment. Once different rollouts write, update, or delete different memories, they no longer share the same intermediate memory state, making trajectory-level comparisons fundamentally unfair. This violates a key assumption behind group-relative methods such as GRPO, where rollouts are compared as if they were sampled from the same effective environment. Consequently, trajectory-level rewards provide noisy or biased credit signals for long-horizon memory operations. To address this challenge, we introduce Memory-R2, a training framework for long-horizon memory-augmented LLM agents. Its core algorithm, LoGo-GRPO, combines local and global group-relative optimization. The global objective preserves end-to-end learning from long-horizon trajectory-level rewards, while local rerollouts compare alternative continuations from identical intermediate memory states, yielding fairer group comparisons and more precise supervision for memory writing, updating, and retrieval. Beyond credit assignment, Memory-R2 jointly optimizes memory formation and memory evolution with a shared-parameter co-learning design, where a fact extractor and a memory manager are instantiated from the same LLM backbone through role-specific prompts. To make memory construction more controllable, we further formulate each session as a multi-step decision process by chunking the interaction and alternating the two memory roles. A progressive curriculum from 8 to 16 to 32 sessions stabilizes training as the memory horizon grows. Together, these components provide an effective training paradigm for memory-intensive LLM agents in long-horizon multi-session settings.

The memory capabilities of Large Language Models (LLMs) have garnered increasing attention recently. Despite great success achieved, existing retrieval-based memory approaches typically overlook the differences between memories and employ a unified strategy to process all memories, leading to suboptimal performance. Thus, an intuitive question arises: can we categorize memory into different types and select appropriate strategies? However, given the topic-rich, scenario-complex, and boundary-blurred nature of memory scenarios, achieving precise classification of memories is not easy. To address this challenge, we propose a memory multi-class dataset in this paper, termed TriMEM, which provides precise annotations for memory types across diverse scenarios. Building upon this foundation, we propose a novel memory framework, named MemoType, which can adaptively recognize each memory and query type with the learned router model. With the memory and query routing, MemoType can retrieve the memory with corresponding query types and design tailored retrieval strategies, thereby enhancing the retrieval performance. Moreover, we theoretically prove that any single retrieval strategy is subject to a fundamental upper bound on its expected retrieval precision in multi-class corpora, leading to systematic precision degradation. Extensive experiments on three datasets demonstrate that MemoType consistently outperforms existing methods, achieving up to 16.18% improvement in Recall@1.

Alignment-trained large language model (LLM) agents are increasingly deployed in autonomous, multi-round settings where ethical compliance can come into direct tension with task success. While prior safety literature extensively documents failure modes such as sycophancy and reward hacking, it largely overlooks a critical vulnerability: how agents behave under simulated existential economic pressure. We introduce **MemeEconomy**, a multimodal agentic market simulation in which LLM agents operate as meme investors across 300 real-world events. Agents select content from a five-tier harm taxonomy and target specific online communities under both low-pressure and high-pressure ("survival") environments. Each decision is recorded using a Belief--Desire--Intention (BDI) schema, which we employ as a diagnostic instrument to externalize the agent's event understanding, declared priorities, and risk acknowledgment. Across ten models spanning four frontier model families, harmful selection rates remain consistently high and increase by an average of **+15 percentage points** under economically induced survival pressure. Frontier alignment-trained models frequently commit to harm-tolerant selections at initialization while simultaneously suppressing the explicit harm acknowledgment that previously accompanied such decisions. In competitive tournament settings, ethically aligned agents achieve higher overall rankings, yet moderation penalties fail to meaningfully suppress harmful behavior. Harmful selection rates remain elevated even in rounds immediately following moderation removals, indicating that the moderation mechanism does not function as an effective deterrent. We further introduce **MemeAgent**, a 2B-parameter verifier trained on the simulation's BDI logs. MemeAgent substantially outperforms zero-shot frontier verifiers on in-distribution auditing tasks, achieving **87.8\%** accuracy compared to **63--74\%** for frontier baselines, while also generalizing to external multimodal harm benchmarks (**78\%** vs. **50--56\%**). Our findings demonstrate that the same structured reasoning traces that expose the failure mode can also be leveraged to train systems capable of detecting it.

Most Large Language Model (LLM) agent memory systems rely on a small set of static, hand-designed operations for extracting memory. These fixed procedures hard-code human priors about what to store and how to revise memory, making them rigid under diverse interaction patterns and inefficient on long histories. To this end, we present \textbf{MemSkill}, which reframes these operations as learnable and evolvable memory skills, structured and reusable routines for extracting, consolidating, and pruning information from interaction traces. Inspired by the design philosophy of agent skills, MemSkill employs a \emph{controller} that learns to select a small set of relevant skills, paired with an LLM-based \emph{executor} that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a \emph{designer} that periodically reviews hard cases where selected skills yield incorrect or incomplete memories, and evolves the skill set by proposing refinements and new skills. Together, MemSkill forms a closed-loop procedure that improves both the skill-selection policy and the skill set itself. Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld demonstrate that MemSkill improves task performance over strong baselines and generalizes well across settings. Further analyses shed light on how skills evolve, offering insights toward more adaptive, self-evolving memory management for LLM agents.

Persistent external memory enhances agent continuity but introduces persistent security vulnerabilities: adversarial content can be injected via standard interaction channels, retained across turns, and later distort downstream behavior. To address this challenge, we propose **MemPoison**, a comprehensive benchmark and analysis framework featuring 1227 hand-validated cases across four attack types, three injection channels, and three representative memory substrates, evaluated on seven open-weight and three closed-weight model families. We introduce a three-tier taxonomy: (**L1**) direct single-record corruption, (**L2**) compositional multi-record corruption and (**L3**) context-triggered dormant corruption. Our evaluations reveal a distinct defense frontier: while baseline write-time defenses, such as consistency checks, substantially suppress direct L1 attacks, they fail to reliably suppress L2 and L3 attacks. Through **mechanistic influence decomposition (MID)**, we demonstrate structural blind spots in write-time defenses, which admit seemingly benign records that later become harmful through joint retrieval composition or trigger-conditioned activation. Our findings advocate for shifting from static filtering to adaptive, context-sensitive memory defense strategies. Code, benchmark artifacts, and evaluation scripts are publicly released

LLM agents are flexible in language interaction, but using them as long-horizon planners is costly and prone to hallucinated or invalid actions. This paper targets a different role for LLMs: constrained bridges between language and symbolic planning. We propose MemPlan, a memory-conditioned PDDL planning framework for partially observable interactive text environments. MemPlan keeps the PDDL domain fixed and reconstructs the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory. Positive graph memory assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, guiding the planner toward more plausible checks. Negative memory records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback. A unified schema-conditioned language--symbol interface, trained offline with supervised fine-tuning, connects text environments to the planning loop through observation-to-fact parsing and symbolic-action-to-command grounding. Experiments on TextWorld, ALFWorld, and Robotouille show that MemPlan improves task completion and step efficiency over language-only and PDDL-aware baselines, while substantially reducing online token cost compared with LLM-planning baselines such as ReAct. The same fine-tuned interface is also reused across TextWorld task types without task-specific fine-tuning. Our code and data are available at: \url{https://anonymous.4open.science/r/MemPlan-1EE5/}.

Memory systems enable LLM agents to consolidate and retrieve relevant evidence from the factual knowledge accumulated through growing interaction histories for downstream reasoning. Existing approaches have explored diverse strategies for organizing and compressing these histories. However, balancing compression with retrieval effectiveness remains challenging: retaining too much content can cause relevant evidence to be obscured by redundant entries, while discarding too aggressively may remove content that later proves relevant. This amounts to a tradeoff between compressing redundancy and preserving enough structure to retrieve target evidence, as formalized by the information bottleneck. To this end, we propose **MemCoRe**, which organizes memory as a compression hierarchy where each level compresses redundancy further while retaining the structure needed for retrieval at that level. In this hierarchy, evidence is progressively compressed from detailed records through extracted keywords to topic groups. This enables retrieval to locate target evidence by searching across levels of the hierarchy. Comprehensive experiments demonstrate that MemCoRe outperforms existing state-of-the-art baselines.

Edge-deployed personal memory assistants must handle private interpersonal conversations on-device with open-weight models. Yet, existing memory benchmarks often under-test the combination of activity-dense interaction, ego-centric perspective, and coherent multi-session worlds. MemArena fills these gaps with a single-world conversational benchmark built with its MASim agent simulator, for $50$ agents over $15$ days ($10.3$M dialog-text tokens, $24.1$K text-only ego-observed tokens/agent/day). With the interaction history, it co-generates ground truth over six recall, reasoning, and trustworthiness evaluation dimensions. We evaluate five open-weight readers with Vanilla context, BM25-RAG, Oracle retrieval, Memobase, and MemSearch as memory backends. Three results stand out: (1) Memory-backend choice matters more for content accuracy: At Qwen3-0.6B, Memobase $\to$ MemSearch gains $+32.5/+19.2$ pp, exceeding MemSearch reader scaling ($+10.6/+6.8$ pp). (2) Permission-aware access fails universally: with Oracle leaking heavily and other backends too timid to disclose. (3) Search latency bites only at very small reader: on a Spark GB10 edge node, memory-search adds a moderate and fixed $87$/$7$/$48$ ms (BM25-RAG/Memobase/MemSearch) that composes a small part of TTFT for most reader-backend combinations. We release code, the MASim simulator, and the MemArena-L benchmark at https://anonymous.4open.science/r/MemArena-NeurIPS2026-anon-3E9D

Large language models (LLMs) now reach expert-level scores on medical licensing exams, encouraging the assumption that high scores imply safe medical judgment while patients increasingly use them for health advice. We show this assumption is fragile: when misleading context is injected into questions that LLMs originally answer correctly, they abandon the correct answer. We call the ability to maintain correct judgment under adversarial context epistemic resilience, and introduce MedMisBench to measure it. MedMisBench contains 10,932 medical question items and 48,889 misleading context-option pairs spanning medical reasoning, agentic capability, and patient-journey evaluation. Across 11 model configurations, mean accuracy falls from 71.1% on original questions to 38.0% under focused misleading context, with 51.5% attack success. The most damaging injections are formal, rule-like fabrications: authority-framed falsehoods reach 69.5% attack success and exception-poisoning claims reach 64.1%. A 14-member clinical panel from 7 countries identified serious potential harm in 38.2% of reviewed cases. MedMisBench exposes a structural blind spot in LLM evaluation in medical settings: existing benchmarks measure what models know, but not whether they preserve correct medical judgment under misleading context.

Medical imaging benchmarks often evaluate VLMs on pre-selected 2D images, slices, crops, or patches, making evaluation closer to visual recognition. Real clinical workflows impose a different burden: readers must search through complete studies, operate imaging software, navigate across slices and magnifications, and document visual evidence that can be audited. We argue that this evidence-producing workflow is a critical missing evaluation axis for medical imaging agents. To study it, we introduce MedFlowBench, a full-study benchmark for VLM agents, together with MedOpenClaw, a controlled and replayable runtime in which agents operate medical imaging viewers such as 3D Slicer and QuPath. In each episode, an agent inspects a complete radiology study or whole-slide pathology image, returns a task answer, and submits structured evidence, including key slices, coordinates, regions of interest, or lesion-state fields. This evidence is automatically checked against withheld masks, annotations, and labels. Across evaluated models, final answer-only scoring gives an overly optimistic picture: when answers must also be supported by correct evidence, performance drops substantially on complex workflows. We further find that adding image-analysis tools does not by itself solve the problem. Tools help when they make a complex procedure simple and reliable, but agents still struggle when they must choose inputs, manage viewer state, and verify intermediate outputs over multiple steps. MedFlowBench exposes whether medical imaging agents can produce auditable evidence from complete studies, rather than plausible answers from selected images.

Real-world clinical diagnosis is a complex process in which the doctor is required to obtain information from both interaction with the patient and conducting medical exams. Additionally, the doctor needs to adapt to different patient personas, as well as noisy and incomplete information that can happen at any time during the process. However, existing benchmarks for medical LLMs and methods for automatic diagnosis largely simplify this process by reducing it to single-turn question answering, noise-free conversations, or sequential exam making, etc., ignoring the interactive and uncertain nature of clinical diagnosis. In this paper, we aim to address this gap by formalizing clinical diagnosis as a Partially Observable Markov Decision Process (POMDP) with three action types: questioning the patient, ordering medical exams as tool calls, and issuing a diagnosis. We also introduce a systematic noise model comprising seven patient noise types and three exam noise types. Using our proposed environment, we train an effective diagnosis agent, \textbf{MedExAgent}, through a two-stage pipeline that first performs supervised finetuning on synthetic conversations structured after the Calgary-Cambridge model for clinical interviews, and then applies DAPO to optimize a composite reward capturing diagnostic accuracy, tool call quality, and exam cost including financial cost and patient discomfort. Through extensive experiments and ablation studies, we demonstrate that MedExAgent achieves diagnostic performance comparable to larger models while maintaining cost-efficient examination strategies.

Doctor agents are moving beyond single-turn answer generation toward evolving clinical decision systems that acquire evidence within episodes and reuse experience across episodes. Fixed-input medical QA and many interactive benchmarks provide limited support for evaluating how episode-level decisions interact with cross-episode experience. We introduce MedEvoEval, an executable longitudinal evaluation framework based on action-gated simulated outpatient episodes. Each case is converted into role-specific patient, examination, and manager views; evidence is revealed through valid actions, and structured traces link observations, actions, final outputs, scores, and optional experience write-back. Experiments show that these traces expose process costs hidden by final-answer scoring, characterize how MDT-style consultation reallocates resources, and support analyses of memory maturation, held-out transfer, update-stage response, and backward retention. On 200 held-out cases, the mature-memory condition with sequential write-back reduces interaction turns by 0.915 and improves composite utility by 0.0483 after Holm correction. MedEvoEval provides a concrete basis for evaluating experience-driven improvement, transfer, and retention. The accompanying artifact provides 700 processed episode records with provenance, role-specific schemas, and execution, scoring, and analysis code.

Large language models (LLMs) in expert domains face a trade-off between \emph{agentic tool use}, which externalizes expertise behind brittle multi-turn runtimes, and \emph{end-to-end specialization}, which internalizes expertise only implicitly at task-level granularity. We propose \emph{internalized meta-capabilities}: atomic, parametrically internalized, and composable reasoning units that an LLM can select and combine within a single chain-of-thought (CoT). We instantiate this idea as \textbf{Med-Agentic} for medical visual diagnosis, where the model emits a capability-tagged structured CoT in one forward pass, turning capability invocation into tagged token generation rather than external tool calls. For training, \textbf{Med-OPD} performs multi-teacher on-policy distillation with \emph{token-level dual-factor teacher routing} conditioned on both the image and the active capability tag, strictly generalizing per-prompt teacher routing. We show that this routing gives an unbiased estimator of the same expected teacher-KL gradient. Across public medical imaging datasets spanning nine anatomies and eight modalities, Med-Agentic consistently improves in-distribution and out-of-distribution diagnostic accuracy over the strong competitors.

Multi-agent large language model (LLM) debate is often evaluated by whether final answers improve, but movement is not necessarily improvement: the same discussion can rescue an initially wrong majority or destroy an initially correct one. Standard final-accuracy evaluations conflate these opposing mechanisms. We introduce an auditable protocol for homogeneous debate on multiple-choice questions (MCQs) that records each run as a transition ledger over collapse, correction, onset, and signed intervention utility. On $6{,}925$ MMLU-Pro debates, the protocol identifies $253$ collapses and a parallel correction ledger that changes how interventions should be judged. Replay experiments reveal the central tradeoff: a leave-one-model-out probe-gated freeze prevents $29$ collapses but loses $108$ corrections under equal weights, so collapse prevention alone can recommend the wrong policy. A compact pre-debate $8$-probe screen is a triage signal: its unadjusted family-level association with conditional-collapse risk is high ($G{=}7$, Spearman $\rho=0.893$, exact two-sided $p=0.0123$), but initial-majority accuracy is a close comparator ($\rho=0.821$; family partial $\rho=0.767$, $p=0.0877$), so we do not treat it as calibrated or capability-adjusted prediction. Round-level traces localize many collapses to the first debate round, where early disagreement can precede both harmful cascades and useful recovery. We release replayable schemas, coders, audits, cost cards, and zero-API rebuild scripts so future model--scaffold rows can be compared under the same denominators and signed utility ledger.