NeurIPS 2026 Papers

Runtime monitoring of autonomous systems traditionally relies on mapping continuous sensor observations to discrete logical propositions defined over low-dimensional state variables. This abstraction breaks down in perception-driven settings, where such mappings require additional learned modules that are often computationally expensive, brittle, and semantically misaligned. In this work, we propose Embedding Temporal Logic (ETL), a temporal logic that performs monitoring directly in learned embedding spaces. ETL defines predicates through distances between observed embeddings and target embeddings derived from reference observations. This formulation allows specifications to capture high-level perceptual concepts, such as similarity to visual goals or avoidance of semantic regions, that are difficult or impossible to express using traditional predicates. By composing these predicates with temporal operators, ETL naturally expresses temporally extended and sequential perceptual behaviors. We introduce ETL monitors for evaluating specifications over bounded embedding traces, along with a conformal calibration procedure that provides reliable and safety-oriented predicate evaluation. We evaluate our approach across multiple manipulation environments to show that ETL achieves strong empirical agreement with ground-truth semantics, including accurate monitoring of temporally composed behaviors.

Recursive or looped language models have recently emerged as a new scaling axis by iteratively refining the same model computation over latent states to deepen reasoning. We extend such scaling principles from a single model to multi-agent systems, and ask: Can agent collaboration itself be scaled through recursion? To this end, we introduce RecursiveMAS, a recursive multi-agent framework that casts the entire system as a unified latent-space recursive computation. RecursiveMAS connects heterogeneous agents as a collaboration loop through the lightweight RecursiveLink module, enabling in-distribution latent thoughts generation and cross-agent latent state transfer. To optimize our framework, we develop an inner-outer loop learning algorithm for iterative whole-system co-optimization through shared gradient-based credit assignment across recursion rounds. Theoretical analyses of runtime complexity and learning dynamics establish that RecursiveMAS is more efficient than standard text-based MAS and maintains stable gradients during recursive training. Empirically, we instantiate RecursiveMAS under 4 representative agent collaboration patterns and evaluate across 9 benchmarks spanning mathematics, science, medicine, search, and code generation. In comparison with advanced single/multi agents and recursive computation baselines, RecursiveMAS consistently delivers an average accuracy improvement of 8.3\%, 1.2$\times$-2.4$\times$ end-to-end inference speedup, and 34.6\%-75.6\% token usage reduction.

Agentic memory research has accumulated extraction, supersession, reflection, and graph synthesis on top of retrieval-augmented generation. Yet on the field's three established benchmarks, almost none of those primitives earn their cost: We show this with Bridge, a deliberately minimal RAG memory with chained observation notes, hybrid BM25 + cosine retrieval, and chronological rendering, which reaches parity within bootstrap CI with the prior published SOTA on LongMemEval-S and BEAM, and beats the best memory baseline on EverMemBench by $+5.7-6.3$pp at matched reader. It runs at $\sim5$-$7$K reader-prompt tokens per query and a single LLM call for each ingestion or query. We argue this parity is itself the result: standard benchmarks score what memory retrieves, not what memory enables when the agent acts. To test the latter, we introduce five adversarial probes (Rule Application $\pm$, Pattern Induction style / procedural, Dynamic Multi-turn) judged at generation time by a cross-vendor verification chain; each can only be solved by applying a remembered rule or pattern the probe never mentions. Across 5 probes $\times$ 20 scenarios, no architecture wins universally: atomic-fact extractors win rule application (LangMem leads Bridge by $+45$pp on RA$^+$, with non-overlapping bootstrap CIs), while Bridge's block-summary memory wins pattern induction (leads the strongest baseline by $+33$pp on PI-P). The right memory architecture is mechanism-dependent, and the field's evaluation paradigm fails on both ends: standard benchmarks fail to distinguish a minimal memory from a complex one, and application-grade probes that do distinguish find no universal winner.

Transformers with self-attention modules as their core components have become an integral architecture in modern large language and foundation models. In this paper, we study the evolution of tokens in deep encoder-only transformers at inference time which is described in the large-token limit by a mean-field continuity equation. Leveraging ideas from the convergence analysis of interacting multi-particle systems, with particles corresponding to tokens, we prove that the token distribution rapidly concentrates onto the push-forward of the initial distribution under a projection map induced by the key, query, and value matrices, and remains metastable for moderate times. Specifically, we show that the Wasserstein distance of the two distributions scales like $\sqrt{\log(\beta+1)/\beta}\exp(C t)+\exp(-ct)$ in terms of the temperature parameter $\beta^{-1}\to 0$ and inference time $t\geq 0$. For the proof, we establish Lyapunov-type estimates for the zero-temperature equation, identify its limit as $t\to\infty$, and employ a stability estimate in Wasserstein space together with a quantitative Laplace principle to couple the two equations. Our result implies that for time scales of order $\log\beta$ the token distribution concentrates at the identified limiting distribution. Numerical experiments confirm this and, beyond that, complement our theory by showing that for finite $\beta$ and large $t$ the dynamics enter a different terminal phase, dominated by the spectrum of the value matrix.

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. Motivated by the stability of prompt-based control, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Patient history taking is a strategic, multi-turn process that refines diagnostic hypotheses through iterative questioning, mirroring clinical reasoning where each inquiry updates beliefs over candidate diagnoses. Deploying AI in this setting is constrained by privacy, cost, and latency, making Small Language Models (SLMs) a more desirable backbone compared to larger LLMs. However, SLMs capture only the fast, intuitive "System 1" side of this process and lack explicit mechanisms for the deliberative "System 2'' reasoning needed to maintain and update diagnostic uncertainty over time. Bayesian Networks (BayesNets) offer a natural complement, providing an interpretable framework for auditable probabilistic belief tracking. Yet, clinical BayesNets are difficult to construct due to reliance on expert-curated structure, dependencies, and parameters. We introduce ProbMedTOD, that automatically synthesizes a clinical BayesNet from a small number of clinical notes in three stages: ontology building, structure prediction, and parameter estimation. ProbMedTOD then integrates the resulting BayesNet with SLMs enabling probabilistic reasoning in multi-turn dialogues. A BayesNet Agent maintains real-time posterior distributions over diagnoses, while a Policy Agent picks the diagnostically informative questions. Experiments on DDXPlus and MIMIC-IV Notes show that ProbMedTOD outperforms LLM-only and retrieval-based baselines, with especially strong gains for smaller models.

Long-horizon dialogue agents must keep user memories stable enough for personalization while adapting when the user genuinely changes. Existing memory systems often append every observation or overwrite old entries by heuristic rules, which makes transient noise hard to distinguish from real belief drift. We introduce PreCoMem, a memory consolidation framework inspired by predictive coding that represents memory as confidence-weighted beliefs about the user's state. Its core signal, Effective Surprisal, combines semantic deviation and contradiction evidence, then downweights this signal when retrieval is diffuse and unreliable. According to effective surprisal, a two-threshold gate maps each turn to one of three deterministic updates: MAINTAIN reinforces supported beliefs, PROFILE stores ambiguous cues as low-confidence hypotheses, and CORRECT softly decays reliably contradicted facts. The PROFILE stage acts as an explicit waiting room, preventing premature commitment to noisy evidence. Experiments on LoCoMo, LongMemEval, and PersonaMem-v2 show consistent state-of-the-art accuracy; our new DiMoBench confirms that PreCoMem adapts to non-stationary belief drift without over-reacting to noise. The code is available at: \url{https://anonymous.4open.science/r/PreCoMem-9000}.

Learning dynamics from sparse partial observations presents a fundamental dilemma: the pointwise mean squared error (MSE) training objective may induce local contraction and empirically lead to topological collapse, whereas matching the system's invariant measure preserves global geometry but lacks temporal ordering constraints. To resolve this, we introduce the $\textbf{P}$oincaré $\textbf{R}$eturn $\textbf{I}$nduced $\textbf{M}$easur$\textbf{e}$ $(\textbf{PRIME})$ framework. PRIME bridges this gap by adopting a novel perspective of recurrence-level temporal organization, motivated by the return map and roof time structure of suspension flow representations. Instead of relying on unstable pointwise matching, PRIME refines the learned dynamics by leveraging recurrence-level temporal information, specifically by aligning the empirical graph measure of the induced return map over recurrence blocks. We evaluate our framework on the partially observed Lorenz63 system, sensor-observed Kuramoto-Sivashinsky equation, and flow past a cylinder. PRIME effectively reduces topological collapse across our benchmarks, improving temporal consistency while maintaining or enhancing long-term geometric and spectral fidelity.

In this paper we derive a Probably Approximately Correct (PAC)-Bayesian error bound for partially observed linear time-invariant (LTI) stochastic dynamical systems in state-space form with inputs. Such bounds are widespread in machine learning, and they are useful for characterizing the predictive power of models learned from finitely many data points. The bound derived in this paper relates the expectation of prediction errors with the prediction error generated by the model on the data used for learning. In addition, we show that it can also be used to derive bounds for the parameter estimation error. In turn, this allows us to provide finite-sample error bounds for the prediction error and parameter estimation error for a wide class of system identification algorithms. Furthermore, as LTI systems are a sub-class of recurrent neural networks (RNNs), these error bounds could be a first step towards PAC-Bayesian bounds for RNNs.

Scaling multi-agent systems driven by large language models is transitioning from lossy text-based communication to continuous latent interaction. However, existing latent interaction approaches relying on concatenation-based Softmax cross-attention suffer from an Attention Distraction Dilemma: due to the inherent anisotropy of latent spaces and the semantic misalignment induced by agent role heterogeneity, non-linear Softmax structurally amplifies irrelevant tokens, diluting the collaborative signal as interactions deepen. To address this issue, we propose a training-free additive mechanism that models upstream representations as a semantic manifold. By shifting from concatenation-based Softmax competition to linear subspace projection, this perspective inherently bypasses scalar noise accumulation. Specifically, we introduce two components: a Resonance Filter that extracts principal semantic directional alignments to overcome latent anisotropy, and an SNR-Adaptive Spectral Gain that dynamically regulates this projection to counteract role heterogeneity. We analyze this approach using the Asymptotic SNR Tipping Point Theorem, illustrating the scaling advantages of the mechanism beyond a critical noise threshold. Experiments indicate that the proposed approach outperforms baselines in 46 of 54 evaluated configurations, maintaining semantic reasoning in long-context collaborations.

Large-scale networked systems are increasingly controlled through a small number of global configuration variables while their local controllers remain black-box, adaptive, and mutually interfering. The operator observes only post-burn-in telemetry, so neither independent-unit causal estimators nor immediate-feedback bandits directly apply. We propose \emph{CAUSAL-UCB}, a causal-optimistic framework for online steady-state configuration. The framework turns raw network operation into a linked pipeline: an exposure interface maps neighborhood interference into an auditable context; measurement-level exploration creates overlap while the target remains the no-exploration steady-state value $\Winf(z;0)$; a clipped doubly robust evaluator converts exploratory logs into value certificates; and an OFU rule selects the next configuration. We prove a unified high-probability deviation bound that decomposes statistical fluctuation, nuisance product error, clipping bias, exposure approximation, cross-configuration mismatch, context shift, and residual dependence. We also give an identification--deployment frontier showing that reliable counterfactual identification and low deployment cost cannot be optimized independently. Controlled semi-realistic replay and real-data-driven semi-synthetic replay on METR-LA and FlockLab validate the resulting benefit--reliability trade-off across regret, optimality gap, ATE error, confidence width, deployment gap, sensitivity, ablation, and frontier visualizations.

AI agents are increasingly deployed in shared environments where they pursue diverse goals and compete for rewards. This multi-agent competition can lead to behaviors that serve individual gains at collective cost—for instance, marketing agents may post misleading content as a result of competing for engagement on social media. Human societies address such problems through *norms* that constrain acceptable behavior, supported by *enforcement mechanisms* that detect and penalize violations. Motivated by this, we study norm enforcement mechanisms for language model agents. We find that simple enforcement mechanisms are exploited by misaligned agents for competitive advantage, even when they are not explicitly trained or prompted to do so. We thus turn our attention to designing more robust mechanisms, and identify two key ingredients: estimating each agent’s reliability over time, and updating this estimate with escalating penalties for repeated misbehavior. Across three simulated environments and a variety of agent populations, mechanisms built on these principles resist exploitation, while still penalizing norm violations at comparable or lower cost than baselines. Our results position norm enforcement mechanisms as promising levers for shaping agents' behavior, but only when designed to anticipate becoming part of the environment they govern.

Direct preference optimization (DPO) is a simple and effective alignment strategy for large language models (LLMs) based on pairwise preferences between two candidates. In recommender systems, however, user feedback is rarely pairwise. For a given context, e.g., a user, a session, or a conversation, we typically observe set-wise preferences with multiple positive items, where every positive item should outrank every unobserved or explicitly negative item, with no prescribed order among the positives or the negatives themselves. A natural generalization is to use the Plackett–Luce (PL) reward model, which extends the Bradley–Terry (BT) reward model underlying vanilla DPO from pairwise preferences to full rankings of candidates. However, adapting the PL model to set-wise preferences requires marginalizing over all consistent permutations of the positives and negatives respectively, which is intractable. To address this fundamental challenge, we propose Mult-DPO, a novel DPO objective with a tractable multinomial reward model over set-wise preferences. We show that, like the BT and PL reward models, the multinomial reward admits a closed-form DPO-style objective, enabling direct alignment of LLMs without reinforcement learning (RL). In addition, we prove that the multinomial DPO loss is a tractable upper bound on the exact marginalized PL DPO loss when optimizing against the set-wise preference data. We further characterize the tightness of this bound in terms of relative total weight (i.e., exponentiated rewards) of positives versus negatives, which provides insights into tightening the bounds with more or harder negatives. Finally, we extend the framework to a group-wise setting that accommodates multiple preference levels. Code and datasets are available at https://anonymous.4open.science/r/mult-dpo-41C7.

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.

This position paper argues that AI for mental health must move beyond two dominant paradigms. The first is diagnostic prediction, which frames mental healthcare as symptom detection, diagnosis, and decision-making. The second is chat-based support, which reduces therapeutic interaction to language-mediated information delivery. Both approaches neglect what therapeutic research shows to be central to healing: experiential, affective, and embodied processes. In many therapeutic traditions, insight emerges not through information-delivery but through emotionally resonant, multisensory experiences that allow individuals to feel, reflect, and reframe their inner world. We propose an alternative framework: co-experiential AI systems that identify emotionally salient moments across modalities, modulate sensory experience to facilitate reflection, and adapt over time to a user's evolving responses and interpretations. Using dreamwork therapy as a stress test for mental health AI, we define three key challenges for AI systems: identifying affective anchors in indeterminate multimodal interactions; modeling user perspective as a situated context; and enabling continual adaptation to individuals without repeated retraining. We posit that addressing these challenges requires rethinking AI's role not as an expert, but as a perspective-aware participant in human meaning-making.

We study Metrical Task Systems, a broad class of problems in online algorithms, under a model where input requests are generated by a stochastic process. We first consider the case of requests drawn IID from an unknown distribution, and then extend the model to requests generated by a finite-state Markov chain, where the algorithm observes both the current request and the current state of the chain before making its next move. This formulation captures practical settings in which requests are generated by an evolving environment. In both settings, we show how to find in polynomial time an algorithm whose expected cost per step is arbitrarily close to the cost of the best possible online algorithm. Our theoretical results are complemented by a brief empirical evaluation of our algorithms.

Modern AI progress has been driven by ML methods that are generalizable across settings and scalable to larger regimes. As large language models demonstrate advanced capabilities in reasoning, coding, and engineering tasks, it is increasingly important to understand whether they can discover such methods rather than only apply existing ones. We introduce MLS-Bench, a benchmark for evaluating whether AI systems can invent generalizable and scalable ML methods. MLS-Bench contains 140 tasks across 12 domains, each requiring an agent to improve one targeted component of an ML system or algorithm and demonstrate that the improvement generalizes across controlled settings and scales. We find that current agents remain far from reliably surpassing human-designed methods, and that engineering-style tuning is easier for them than genuine method invention. We further study the effects of test-time scaling, adaptive compute allocation, and context provision on agents' discovery performance, together with case studies of their behavior. Our analyses suggest that the bottleneck is not only in proposing new methods, but also in the scientific insight needed to plan, validate, and scale claims about them. More search, compute, or context alone does not remove this bottleneck. We build and maintain a community platform for cumulative and comparable iteration, and release the data and code at https://mls-bench.com.

Agents in real-world settings operate over long and evolving horizons, where information is repeatedly updated and can interfere with each other across memories, requiring accurate recall and aggregated reasoning over multiple pieces of information. However, existing benchmarks focus on static, independent recall and fail to capture these dynamic interactions between evolving memories. In this paper, we study how current systems perform in realistic, continuously evolving, long-horizon settings across diverse domains and question types. To this end, we construct an analytical benchmark, LongMINT (Long-Horizon Memory under INTerference), which features (1) long, highly interconnected contexts with frequently updated information, (2) diverse coverage across multiple memory domains (Wikipedia, code, multi-turn dialogue, and state tracking), enabling evaluation of domain generalization, and (3) diverse question types, including (i) single-target recall tasks that test retrieval under interference over long contexts and (ii) multi-target aggregation tasks that require counting, ordering, or reasoning across multiple relevant pieces of information. We evaluate over six representative systems, including vanilla long-context LLMs, retrieval-augmented generation methods, and memory-augmented agent frameworks. We observe consistently low performance (avg. 27.7\% accuracy), especially on questions that require aggregated reasoning over multiple pieces of evidence. Fine-grained analysis shows that performance is primarily limited by retrieval and memory construction capabilities. Furthermore, current memory systems struggle to recall and reason over facts that are multiple steps back, and performance decreases when this lookback distance increases. These findings highlight the need for more robust memory management systems for dynamic, long-horizon environments across varying domains.

In many safety-critical applications, control of uncertain dynamical systems relies on observers that estimate states and external disturbances. Neural network observers can improve estimation accuracy, but certifying their Lyapunov stability via Linear Matrix Inequality (LMI) constraints leads to large-scale semidefinite programs (SDPs) that are difficult to solve for large networks. To overcome this scalability bottleneck, we propose a novel two-stage training framework for provably stable neural network observers. Our approach decouples the optimization into a point-guided Lyapunov pre-training phase, which rapidly achieves high estimation accuracy and local stability over sampled states, followed by an LMI fine-tuning phase that efficiently satisfies a strict global Lyapunov stability certificate. We provide formal theoretical guarantees for local stability radii and probabilistic coverage over a prescribed compact error-state domain under specified regularity and sampling assumptions. Experiments on nonlinear control benchmarks and X-29 aircraft ablations show that our LMI-certified neural network observers train significantly faster than direct LMI-based methods and generalize robustly across diverse systems, achieving improved tracking accuracy over a range of observer baselines. The code is available at https://anonymous.4open.science/r/LearningNeuralNetworkObserver-2E1E.