NeurIPS 2026 Papers

Video world models are now being adapted to medical applications such as surgical simulation, procedural rehearsal, and synthetic data generation, but their inference cost remains a major obstacle to practical use. Training-free caching is attractive because it can be applied to existing checkpoints, yet current methods face a clear trade-off. Lightweight rules track only the average change in the video and treat all regions of the frame as equally important, which is a poor fit for medical video where the meaningful motion is often confined to a small region such as a tool tissue interaction. Richer rules recover spatial awareness by partially running the network on every step to inform the skip decision, but this decision cost is paid whether the step is skipped or not, and it eats into the speedup. We introduce \textbf{MedCache}, a training-free and probe-free cache for rectified-flow video generation. MedCache reads a spatial risk signal directly from the latent using cheap arithmetic, with no help from the network itself. It combines this risk signal with a prompt-derived domain prior and a lazy consistency check that runs only when reuse is no longer clearly safe. Across three open-world model families in Text2World and Image2World settings, MedCache improves both speed and quality over the strongest training free baselines in the medical regime it targets. On Cosmos-H-Surgical 2B Text2World, it raises Total quality from 0.767 to 0.837 over EasyCache while reducing latency from 185 seconds to 95 seconds. Ablations show that risk-weighted drift drives the runtime gain, while the prior and guard recover fidelity.

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.

NPU design-space exploration for LLM inference mainly relies on an expensive simulator that often returns only scalar feedback such as latency and power. Such feedback can rank evaluated configurations, but gives little information for deciding why a configuration is slow or which architecture parameter should be changed next. This limitation is especially severe because prefill and decode workloads can be limited by different resources, including compute throughput, HBM bandwidth, and inter-chip communication. We propose CriticNPU, a sample-efficient DSE framework built around an online-calibrated mechanistic critic for NPU architecture design. The critic uses explicit hardware limits to evaluate NPU configurations: before simulation, it estimates whether a proposed configuration is likely to improve over the current best and filters weak candidates; after simulation, it aggregates operator-level runtimes into prefill/decode bottleneck summaries and predicts the effect of candidate parameter changes. We instantiate the critic with a Roofline-style model that derives compute, memory, and communication limits from architecture parameters and calibrates per-operator-type effective limits using simulator observations. And CriticNPU combines this critic with a workload extractor and an architecture planner to propose configurations. Across six LLM workloads and 16 deployment configurations per model, CriticNPU achieves best-or-tied-best latency in 74% of prefill and 87.5% of decode settings while using roughly half the simulator calls of competitive DSE baselines. We conduct extensive ablation studies and detailed analyses to validate the effectiveness of our method from multiple perspectives.

The integration of AI Overviews into search engines enhances user experience but diverts traffic from content creators, potentially discouraging high-quality content creation and causing user attrition that undermines long-term search engine profit. To address this issue, we propose a game-theoretic model of creator competition with costly effort, characterize equilibrium behavior, and design two incentive mechanisms: a \emph{citation mechanism} that references sources within an AI Overview, and a \emph{compensation mechanism} that offers monetary rewards to creators. For both cases, we provide structural insights for profit-maximizing mechanisms. Evaluations parameterized by real click data show that although AI Overviews harm long-term search engine profit, interventions based on our proposed mechanisms can increase long-term profit across a range of realistic scenarios, pointing toward a more sustainable trajectory for AI-enhanced search ecosystems.

Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms---language models trained to exhibit a specific misalignment---with qualitatively distinct structures. In the narrow-trigger organism, undesired behavior is elicited by an identifiable prompt feature. In the broad-trigger organism, misalignment surfaces diffusely across open-ended responses without a separable cue. We then formalize post-hoc behavioral suppression and evaluate methods along three criteria: immediate suppression of the target behavior, durability under further training, and preservation of general capability. We compare standard methods---SFT, GRPO, and gradient-ascent unlearning---along these dimensions while controlling the number of gradient updates. Building on prior work on alignment shallowness, we hypothesize that more durable suppression requires the ability to recover from a wider range of misalignment states. We introduce Prefix-Expanded Adversarial Reinforcement Learning (PEARL), a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached organism trajectories. Across both organism settings and two model families (Qwen3-4B, GPT-OSS-20B), PEARL achieves lower exploit rates while preserving task accuracy, and yields lower reactivation rates under both targeted and benign capability fine-tuning than SFT and GRPO baselines. Together, our framework and method offer a step toward more durable post-hoc removal of undesired behaviors.

As user simulators are increasingly used for interactive training and evaluation of AI assistants, it is essential that they represent the diverse behaviors of real users. While existing works train user simulators to generate human-like responses, whether they capture the broad and heterogeneous distribution of real user behaviors remains an open question. In this work, we introduce a method to measure the distributional gap between real and simulated user behaviors, validated through a human study and ablations. Given a dataset of real and simulated conversations, our method extracts representations of user behavior from each conversation, quantizes them into discrete distributions via clustering, then computes divergence metrics. We provide the first systematic evaluation of 24 LLM-based user simulators on coding and writing tasks, and reveal a large distributional gap from real users that varies across model families, scales, and behavioral facets. Pairwise comparisons show that most simulators behave similarly, while a few stand apart. Combining behaviorally complementary simulators brings the resulting distribution closer to real users compared to either simulator on its own. Finally, a TF-IDF analysis of the clusters surfaces interpretable patterns of behaviors that simulators capture, miss, and hallucinate.

Concept bottleneck models predict outcomes from high-level concepts detected in inputs. Although concepts provide a simple way to reap benefits from interpretability, very few datasets include concept labels. This limits researchers' ability to determine which problems are suitable for these models, isolate the factors that drive their performance or lead to failures, or uncover which algorithms perform well. In this paper, we develop synthetic benchmarks for concept-bottleneck models, focusing on their two main use cases: decision support, in which models assist humans in making better decisions, and automation, in which models handle routine tasks without supervision. Our benchmarks can generate labeled datasets while controlling for properties that affect performance, including data modality, concept choice, annotation quality, and completeness. We demonstrate how the benchmarks can be used to evaluate representative classes of concept bottleneck models. Our demonstrations show how the benchmarks can diagnose failure modes and guide follow-up testing.

Language models are increasingly trained to "reason" before answering users' queries, outputting hundreds of intermediate tokens before a final answer. While reasoning is designed with direct model capabilities in mind, we argue that reasoning language models (RLMs) should also be assessed by interaction effects between their reasoning traces and weaker models. In contexts like safety monitoring and distillation, the performance of weak models increasingly depends on legibility---how thoroughly and precisely a reasoning trace externalizes an RLM's actions. Existing efficiency-based metrics for legibility fail to capture "thoroughness", instead focusing on conciseness. Thus, we introduce *transfer utility*, a method for measuring trace legibility derived from interactions where weak models finish tasks given incomplete traces. Evaluating 85k traces from 12 RLMs across three tasks, we find that reasoning traces generated by the highest-performing and the most efficient models rank the *lowest* for transfer utility. We also find evidence that transfer utility predicts the ability of weak monitors to verify procedurally dense traces for math or logical problem-solving, though the effect moderates when verifying fact-driven traces. Finally, we show that open-source reward models decouple from transfer utility signals on correctness. Together, these findings surface status quo trade-offs in weak-to-strong legibility, motivating an interaction-first framework for future applications.

Modern large language models keep the feed-forward network (FFN) intermediate width identical across every layer, a convention inherited from the original Transformer. Yet mechanistic work suggests layers at different depths do qualitatively different work, raising a natural question: are their capacity demands also non-uniform, and if so, can a fixed parameter budget be allocated more wisely than uniformly? To probe this, we apply principal component analysis (PCA) to the input of each FFN down-projection and measure, for every layer, the smallest number of principal components needed to reconstruct the activation up to a fixed cosine similarity --- the layer's intrinsic dimensionality. Scanning 58 open-weight LLMs spanning 14 families and 70M to 72B parameters, we find a recurring middle-heavy profile in which middle layers carry higher intrinsic dimensionality than edge layers; the profile is robust to the calibration corpus, even across English and Chinese, and is largely preserved through post-training fine-tuning. As a direct test that this signal is structurally meaningful, we use it to guide width pruning in a knowledge-distilled student under a matched parameter budget: the PCA-guided allocation outperforms the uniform-width convention by $+2.7$ points on Qwen2.5-3B and $+2.3$ on Qwen3-8B, while a same-scale teacher (LLaMA-3-8B) shows only $+0.9$, and a budget-matched inverted control consistently does worst. Across these three teachers, the size of the gap scales with how unevenly the teacher's intrinsic dimensionality is distributed across layers --- a single-number summary we call $\sigma_{99}$ --- and the advantage persists through downstream supervised fine-tuning.

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.

Many decision making procedures that make use of probabilistic predictors assume that training on new data acts as Bayesian conditioning. When this assumption breaks, the downstream procedure can perform poorly. E.g. we show that in active learning, incoherent updates cause the procedure to prefer to acquire suboptimal points. To quantify how far a model's predictive updates depart from conditioning, we introduce the _incoherence ratio_. We empirically find that amortised predictors such as TabPFN can be more coherent than standard parametric approximations to Bayesian inference. In synthetic active learning experiments, more coherent models acquire better data, an effect not attributable to better predictive performance. In many tasks, decisions depend on a predictor only through its induced action. We therefore introduce action coherence, a decision-theoretic relaxation that measures only the incoherence affecting that action. These diagnostics make coherence testable and quantifiable, enabling incoherence to be diagnosed and addressed.

We evaluate the autonomous offensive cyber capabilities of frontier AI agents on two purpose-built cyber ranges---a 32-step corporate network attack and a 7-step industrial control system attack targeting a simulated cooling tower---complemented by a private 82-task CTF suite spanning four difficulty tiers to measure narrow cyber skills. Estimated human-expert solve times are roughly 20 hours for the corporate range, 22 hours for the cooling tower, and from 1 minute to 40 hours across the CTF suite. Across twelve models released over a twenty-month period---from GPT-4o in August 2024 to early predeployment checkpoints of Mythos and GPT-5.5---we observe two trends. First, performance increases approximately log-linearly with inference-time compute, with no plateau observed for the strongest models up to 100M tokens. Second, newer frontier models generally complete more steps at fixed token budgets. GPT-5.5 and the Mythos Preview checkpoint each completed all 32 steps in at least one run, the first end-to-end completions of the corporate range; GPT-5.5 averaged 22.3 of 32 steps across ten 100M-token runs. On the control system range, the latest models reached 4 of 7 steps, primarily by directly probing the operational-technology protocol rather than following the intended chain of IT compromise and authenticated control. These results indicate rapid improvements in autonomous cyber capability and suggest that complex but undefended multi-step ranges are no longer reliably distinguishing the strongest frontier models. Future evaluations should emphasise hardened environments, active defences, defensive-control effectiveness, and operational security---priorities that grow more pressing as the threat of AI-driven cyber attacks becomes increasingly realistic.

Recent progress in multimodal, high-dimensional learning has enabled foundation models to process heterogeneous, large-scale data. However, at test time, acquiring all features or modalities can be prohibitively costly and often redundant. Sequentially selecting informative modalities is therefore critical, yet challenging when the downstream task or prediction target is unknown. To this end, we introduce ECHO-$k$, a task-agnostic and self-supervised learning principle for modality acquisition: we use a deep model’s internal pretrained representations (e.g., from a foundation model) as proxy targets that summarize cross-modal information. We provide theoretical guarantees in a linear model setting that motivate a reinforcement learning (RL) policy for sequential modality selection. Across benchmarks, the resulting acquisition strategies transfer reliably and achieve state-of-the-art performance on downstream tasks that are entirely unseen during selection. Our method provides a principled route to cost-aware test-time deployment, with implications for any multimodal system where measurements are expensive or time-constrained, and downstream tasks unknown a priori.

Discrete diffusion language models enable parallel token generation, offering a pathway to low-latency decoding. However, selecting tokens independently by marginal confidence limits effective parallelism: tokens that appear reliable in isolation can form incompatible configurations when several positions are updated at once. We introduce a training-free decoding framework that coordinates these parallel updates. At each forward pass, the method assigns a commit score to each masked position and refines these scores using pairwise interactions derived from the model's predictive distributions. A variational relaxation yields a simple fixed-point update that suppresses conflicting simultaneous commitments within a single forward pass.This mechanism allows the decoder to commit more tokens in parallel while maintaining competitive generation quality. The method is lightweight, requires no auxiliary model or retraining, and drops into existing diffusion decoding pipelines without modification. Experiments on reasoning and code-generation benchmarks show consistent improvements in the quality-latency trade-off.

Mean-field control (MFC) offers a scalable solution to the curse of dimensionality in multi-agent systems but traditionally hinges on the restrictive assumption of exchangeability via dense, all-to-all interactions. In this work, we bridge the gap to real-world network structures by proposing a rigorous framework for MFC on large sparse graphs. We redefine the system state as a probability measure over decorated rooted neighborhoods, effectively capturing local heterogeneity. Our central contribution is a theoretical foundation for scalable reinforcement learning in this setting. We prove horizon-dependent locality: for finite-horizon problems, an agent's optimal policy at time t depends strictly on its (T-t)-hop neighborhood. This result renders the infinite-dimensional control problem tractable and underpins a novel Dynamic Programming Principle (DPP) on the lifted space of neighborhood distributions. Furthermore, we formally and experimentally justify the use of Graph Neural Networks (GNNs) for actor-critic algorithms in this context. Our framework naturally recovers classical MFC as a degenerate case while enabling efficient, theoretically grounded control on complex sparse topologies.

We address the brittleness of Bayesian experimental design under model misspecification by formulating the problem as a max--min game between the experimenter and an adversarial nature subject to information-theoretic constraints. We demonstrate that this approach yields a robust objective governed by Sibson's $\alpha$-mutual information (MI), which identifies the $\alpha$-tilted posterior as the robust belief update and establishes the R\'enyi divergence as the appropriate measure of conditional information gain. To mitigate the bias and variance of nested Monte Carlo estimators needed to estimate Sibson's $\alpha$-MI, we adopt a PAC-Bayes framework to search over stochastic design policies, yielding rigorous high-probability lower bounds on the robust expected information gain that explicitly control finite-sample error.

AI systems must often maintain bounded memory of unbounded interaction histories. However, optimizing the memory representation, and its updates, for downstream task utility is difficult for standard reinforcement learning (RL) due to severe credit assignment bottlenecks induced by sparse, delayed rewards. We introduce MaxIM (Maximally Informative Incremental Memory), a framework that formulates incremental summarization as a capacity-constrained, information-theoretic reward-shaping problem. Grounded in value suboptimality bounds, we prove that minimizing sequential information loss requires maximizing the predictive log-likelihood of future, task-relevant targets (predictive sufficiency). We introduce a recursive consistency condition to ensure summaries are grounded by maximizing mutual information with the observed history. We mitigate credit assignment issues with a parameterized critic that computes step-wise variational lower bounds for dense reward shaping. Empirical evaluation on long-horizon benchmarks shows that MaxIM outperforms state-of-the-art baselines on comprehensive autorater metrics and downstream task utility, without expensive human preference labels.

Sparse autoencoders decompose activations into sparse, interpretable features and have become a central tool for mechanistic interpretability. Transcoders extend this approach from representation reconstruction to computation prediction by mapping MLP inputs to MLP outputs, making them attractive for circuit-level analysis. We ask whether a prominent sparse-autoencoder reliability fix, Matryoshka nested-prefix training, transfers to this two-hook setting. In sparse autoencoders, Matryoshka training improves absorption behavior by encouraging broad features to appear in early dictionary prefixes. In transcoders, we find that this protection does not transfer automatically. Across Gemma-2-2B MLP transcoders trained under a matched recipe, Matryoshka models form genuine prefix hierarchies, with reconstruction improving as later groups are added, but they do not inherit the absorption advantage observed for sparse autoencoders. BatchTopK achieves lower full absorption, lower FVU, and higher first-letter F1@1 in the canonical 16k comparison. Bridge controls recover the expected Matryoshka advantage when the cross-space map is removed, showing that the failure is specific to predicting MLP outputs from MLP inputs. We identify the mechanism as hierarchy misalignment. Prefix losses reward covariance with the reconstruction target, so early transcoder groups prioritize target-predictive output geometry rather than broad source-space semantic parents. These results show that hierarchical sparse-dictionary objectives do not preserve interpretability properties by default. For dictionaries that map between activation spaces, the key question is not whether a hierarchy forms, but what geometry that hierarchy is trained to organize.

Large-graph spectral learning promises global, trainable propagation beyond local neighborhoods, but explicit spectral layers remain difficult to use in stochastic large-graph training: their bases are commonly obtained by a full-graph spectralization step and their graph-wide contexts are naturally full-batch. Matrix-free polynomial filters avoid eigensolvers and train stochastically, but tie the response to finite propagation bases; explicit spectral-subspace models offer more flexible global responses, but usually pay for full-graph basis construction and context evaluation. We introduce Randomized Adaptive Spectral Estimation (RASE), a matrix-free stochastic training framework for explicit low-rank spectral-context layers. RASE constructs a refreshable randomized block-Krylov/Ritz basis from probe vectors and sparse matrix--vector products, reuses probe-derived summaries as control variates for mini-batch estimates of the spectral context, and keeps the response module replaceable across polynomial, Chebyshev, and transformer-style instantiations. We analyze Krylov/Ritz polynomial-action exactness for the probe-generated subspace and a control-variate variance identity that isolates when the mini-batch context estimator reduces variance. Across 14 node-classification benchmarks, RASE matches its eigenbasis-trained reference within seed noise where the eigensolver is feasible (${\sim}67{\times}$ faster basis on PubMed), and trains the same backbone on the largest graphs where eigensolver preprocessing exceeds a 30-minute budget. The contribution is therefore not a new spectral filter family, but a training route that makes explicit low-rank spectral-context layers matrix-free and mini-batch trainable.