NeurIPS 2026 Papers

Many recent flow-matching and diffusion-style generative models rely on auxiliary stochastic dynamics during training: a richer process is simulated to define conditional targets, but the auxiliary state is either intractable to sample at generation time or simply not part of the desired output. Existing Generator Matching theory formalises conditioning on static latent random variables, and several recent papers prove special cases of projection results for particular augmented-state constructions. We introduce latent process generator matching, a general framework that treats the observed generative state as a deterministic image $X_t=\Phi(Y_t)$ of a tractable Markov process $Y_t$. We show that in this setting one may learn the generator of a stochastic process on the image space which has the same one-time marginal distributions as the projected process. This generalizes and subsumes the discrete latent process results from the literature, and extends Generator Matching from static latent variables to a rich family of time-dependent latent conditional processes.

Video diffusion models often fail in precisely the dimension that distinguishes videos from images: motion. Standard reconstruction and flow-matching objectives in pixel or VAE-latent space are dominated by appearance, so visually similar clips can be treated as close even when their underlying dynamics differ. We propose Latent Motion Alignment, a training-time objective that compares predicted and ground-truth videos in a learned motion space. This space is learned by training a compact encoder to describe how one latent frame changes into the next, while a decoder ensures that the resulting token captures information needed to model the transition. Once learned, the encoder is frozen and used as an auxiliary loss for fine-tuning video diffusion models. The motion module is used only during training and adds no inference-time cost. We validate that the learned space separates motion rather than appearance shortcuts, improves motion-sensitive generation metrics, and enables fine-grained Olympic-diving generation where jump type and score depend on subtle temporal structure.

We uncover a latent capacity for introspection in a Qwen 32B model, demonstrating that the model can detect when concepts have been injected into its earlier context and identify which concept was injected. While the model denies injection in sampled outputs, logit lens analysis reveals clear detection signals in the residual stream, which are attenuated in the final layers. Furthermore, the effect is heavily dependent on context – e.g., prompting the model with accurate information about AI introspection mechanisms can dramatically strengthen this effect: the sensitivity to injection increases massively (0.3% → 39.9%) with only a 0.6% increase in false positives. Also, mutual information between nine injected and recovered concepts rises from 0.62 bits to 1.05 bits, ruling out generic noise explanations. Our results demonstrate models can have a surprising capacity for introspection and steering awareness that is easy to overlook, with consequences for latent reasoning and safety.

Reliable physics simulation requires both broad generalization across heterogeneous PDE families and stability under long autoregressive rollouts. Current neural solvers rarely provide both. Deterministic operators accumulate rollout error, while probabilistic solvers typically remain tied to a single PDE family or a short prediction horizon. We address this gap with the \textbf{Latent Generative Solver} (LGS), which couples a Physics VAE (PhyVAE) with a Pyramidal Flow-Forcing Transformer (PFlowFT). PhyVAE compresses states from twelve PDE families into a shared \emph{physical manifold}, separating dynamics-relevant structure from the ambient state space. PFlowFT then predicts the next latent state through input-noised flow matching, yielding a sufficient-condition contraction mechanism that explains improved autoregressive rollout stability. Pretrained on a 2.5\,M-trajectory, 16-system corpus at $128^2$ resolution, LGS obtains the lowest relative $L_2$ error (L2RE) on 15 of 16 systems at both 5- and 10-step rollout. At 20 steps, LGS reduces the L2RE from $56.1\%$ to $\mathbf{30.2\%}$ against matched deterministic and adapted generative baselines, while requiring $\mathbf{13}$--$\mathbf{77\times}$ less recurrent dynamics-step compute. On a held-out $256^2$ Kolmogorov flow, LGS reduces the 1-step L2RE from $0.398$ to $0.129$ within five finetuning epochs; U-AFNO achieves only $0.653$ to $0.343$ under the same protocol.

Understanding the internal “thinking” process of Large Language Models (LLMs) and the cause of hallucinations remains a key challenge. To this end, we introduce latent debate, a structured surrogate framework for interpreting model outputs on True/False prediction tasks through the lens of internal latent arguments and interactions amongst them. Unlike human debates, latent debate captures the hidden supporting and attacking signals that arise within a model during a single inference. We first present a model- and task-agnostic conceptual framework, and then instantiate it symbolically to approximate the thinking process of LLMs towards binary decisions. Empirical studies demonstrate that our latent debate is a faithful structured surrogate model that has highly consistent predictions with the original LLM, while providing a form of interpretability. We also demonstrate that our latent debate provides a strong baseline for hallucination detection. Specifically, we identify strong correlations between debate patterns and hallucinations, such as a high degree of disagreement in the middle layers of the latent debate surrogate is linked to a higher risk of hallucinations. Our findings suggest that latent debate shows potential to analyze internal signals in LLMs for binary decision settings.

Retrieval-Augmented Generation (RAG) has become a standard approach for enhancing large language models (LLMs) with external knowledge, mitigating hallucinations, and improving factuality. However, existing systems rely on generating natural language queries at each hop and maintaining a strict architectural separation between retriever and generator, preventing them from leveraging the full representational capacity of the LLM. We propose LAnR (Latent Abstraction for RAG), a unified framework in which a single LLM jointly performs encoding, retrieval, and generation entirely within its own latent space. Rather than generating textual queries, LAnR produces dense retrieval vectors from the hidden states of a designated [PRED] token and uses them to match against encoded document representations from the same model. Furthermore, LAnR adaptively decides when sufficient evidence has been retrieved using a lightweight MLP control head over those same hidden states, eliminating explicit token-level stopping reasoning. Extensive experiments on five QA benchmarks spanning single-hop and multi-hop settings demonstrate that LAnR outperforms existing RAG methods, while achieving improved inference efficiency through reduced number of retrieval calls and tighter model integration.

Asynchronous Partial Rollouts (APR) accelerate RLHF by training on the first $k$ of $N$ parallel completions, but this speed hides a critical flaw: because autoregressive generation time scales linearly with token count, a wall-clock cutoff acts as an invisible length filter. On reasoning tasks where reward correlates positively with trajectory length, APR systematically deletes the most valuable, multi-step reasoning chains from the training batch. We formalize this phenomenon as **Latency-Conditioned Selection Bias** (LCSB). We prove that whenever reward and selection probability negatively correlate ($\mathrm{Cov}_{p_\theta}(R, \alpha) < 0$), the APR objective structurally attenuates and can even invert the on-policy gradient. Crucially, standard importance-sampling corrections (e.g., V-trace) cannot resolve this because all trajectories strictly originate from the behavior policy. After empirically validating this latency-length filter in a deployed vLLM system and demonstrating catastrophic, monotonic performance degradation on reasoning benchmarks, we provide two actionable solutions. We introduce **Effective Gradient Information** (EGI), a diagnostic metric that detects LCSB thousands of steps before reward curves diverge, and **Selection-Probability Weighting** (SPW), a near-zero-overhead correction that mathematically neutralizes the bias to recover on-policy reasoning performance without sacrificing asynchronous throughput.

In this paper, we study last-iterate convergence in stochastic constrained convex--concave minimax optimization. A key difficulty is that the last iterates of vanilla stochastic extragradient (S-EG) and stochastic optimistic gradient descent--ascent (S-OGDA) can fail to converge in the presence of gradient noise, even for simple bilinear problems. To address this issue, we regularize the original convex--concave problem into a strongly convex--strongly concave one. Applying S-EG and S-OGDA to the regularized problem gives two simple single-loop first-order methods, which we call perturbed S-EG and perturbed S-OGDA. By carefully choosing the regularization parameter and balancing the resulting regularization bias, stochastic error, and optimization error, we prove that both methods achieve a last-iterate convergence rate of $\mathcal{O}(T^{-1/4+\varepsilon})$ for any $\varepsilon>0$ in terms of the primal--dual gap. This improves upon the best-known $\tilde{\mathcal{O}}(T^{-1/7})$ rate under comparable constrained settings. Moreover, for unconstrained problems, we establish almost sure convergence under a single-timescale stepsize scheme, whereas prior almost sure convergence results typically rely on two-timescale stepsizes and are limited to S-EG.

The ability to learn a new task from a few examples has been crucial to the success of foundation models such as large language models (LLMs). However, current foundation models for structured data like tables and relational databases still require tens of thousands of labeled examples to perform well on new tasks. Here we show that when pretrained at scale with the right recipe, Relational Transformers (RTs) can make state-of-the-art (SoTA) predictions with only hundreds of labels. This capability arises from the following ingredients: First, we assemble THE JOIN, the largest pretraining corpus of relational data to date, comprising 6k forecasting tasks across 650 real-world databases from diverse domains. Second, we develop a pretraining recipe that combines (1) mixed context sizes for multi-scale learning, (2) multi-cell masking for dense supervision, and (3) a novel random- walk-based retriever to efficiently gather relevant context. Third, we identify context ensembling and context tuning as complementary axes for scaling test-time compute to further improve few-shot performance. Pretrained with our recipe on THE JOIN, RT achieves parity with LLM Agent + TabICLv2 and RDBLearn + TabICLv2 pipelines with 32–23×fewer labels respectively, and even surpasses the prior SoTA of full task-specific training, on average nMAE for RelBench regression tasks. Context ensembling and tuning improves this further by up to 3%, and 4% respectively. Our ablations highlight the importance of schema semantics, multi-cell masking, and random-walk retrieval. Overall, our work paves the way for developing relational foundation models with strong few-shot capabilities.

Covariate-adjusted response-adaptive randomization (CARA) designs can improve statistical efficiency and participant welfare in randomized experiments by learning from accrued data and dynamically updating treatment allocation. In practice, two issues limit their impact. Many experiments are sample-size constrained because enrollment is difficult and trials are expensive. At the same time, studies increasingly collect rich pre-treatment information, including unstructured data such as text, images, and clinical notes, yet most CARA designs rely on a small set of structured covariates when computing allocation probabilities. Recent advances in large-scale pre-trained large language models (LLMs) offer a new opportunity to extract useful signals from unstructured inputs and external knowledge. However, few-shot LLM predictions are not ordinary fixed fitted predictions because prompt demonstrations sampled from accrued trial data create data-dependent randomness and can induce correlation across predictions. If used naively to drive allocation updates, AI-generated signals may weaken CARA's ability to improve power and participant welfare. We propose a CARA design that integrates few-shot LLM-predicted outcomes through resampling-based aggregation, calibration, and effective residual variances to guide both adaptive treatment allocation and downstream treatment effect estimation. Theoretical results and simulation studies show that, when calibrated few-shot LLM auxiliary signals reduce residual variation, coupling an LLM-enhanced estimator with an adaptive allocation rule can improve statistical efficiency and participant welfare.

Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values ranging from 0.83 to 0.99 across HMM and LLM combinations, typically occurring at early-to-middle layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade substantially. Together, these results provide representation-level evidence that ICL in pre-trained LLMs approximates optimal Bayesian prediction over a context-inferred HMM. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale transformers trained on natural language.

Behavior policies are often formulated as continuous generative models, whose iterative denoising processes are expressive but difficult to interpret and prone to producing implausible actions. We propose the Large Discrete Policy (LDiP), a fully discrete behavior modeling framework that selects actions from a large vocabulary of physically plausible candidates. Rather than perturbing actions, LDiP improves expressivity through stochastic iterative scoring: it progressively re-scores and prunes candidates with score-space stochasticity, enabling fine-grained ranking and exploration among plausible actions while preserving an explicit decision process. Across end-to-end planning, closed-loop driving, robotic manipulation, and vision-language-action settings, LDiP consistently outperforms strong discrete and continuous baselines in autonomous driving, and exceeds or matches continuous generative policies in robotic manipulation. These results show that discrete policies, when equipped with effective scoring mechanisms, offer an expressive, plausible, and interpretable alternative for behavior modeling.

Transformers update token representations through multi-head attention and residual connections as $X \leftarrow X + \sum_{i} P^{(i)}XW_{V_i}W_{o_i}$, where $P^{(i)}$ is the softmax attention matrix in head $i$. We propose replacing a subset of $P^{(i)}$'s with the Laplacian $I - P^{(i)}$, giving $X \leftarrow X + \sum_{i \in \mathcal{A}} P^{(i)}XW_{V_i}W_{o_i} + \sum_{i \in \mathcal{L}} (I - P^{(i)})XW_{V_i}W_{o_i}$. Our proposal has two motivations. First, it allows attention heads to update the mean of token representations, while Laplacian heads can directly control within-sequence variance. Second, if tokens are viewed as nodes in a graph with edge weights \(P^{(i)}\), then \(I - P^{(i)}\) is the corresponding graph Laplacian, and the update can be interpreted as one step of heat diffusion on the graph. We show that this simple modification improves performance across supervised learning, language modeling, and self-supervised learning tasks. To investigate why, we examine the token representations learned with and without Laplacian heads. In supervised learning, Laplacian heads collapse token representations within the same sequence and align the sequence means with the geometry of Neural Collapse. In language modeling, they increase the separability of token representations that share the same next-token prediction. In self-supervised learning, they produce token representations whose principal components are better suited for segmentation. Across modalities, they also lead to faster-decaying spectra, indicating stronger token smoothing. Overall, our findings challenge the prevailing view that token oversmoothing is inherently harmful, showing instead that certain forms of smoothing can be beneficial.

Domain adaptation faces a fundamental paradox in the cold-start regime. When target data is scarce, statistical methods fail to distinguish relevant source domains from irrelevant ones, which often leads to negative transfer. In this paper, we address this challenge by leveraging expert textual descriptions of the target domain, a resource that is often available but overlooked. We propose a probabilistic framework that translates these semantic descriptions into a choice model, namely a Language-Induced Prior (LIP), that learns the preferences from a pre-trained Large Language Model (LLM). The LIP is then integrated into an Expectation-Maximization (EM) algorithm to identify source relevance. Methodologically, this framework is compatible with any parametric model where a likelihood is available. It allows the LIP to guide the selection of sources when target signals are weak, while gradually refining these choices as samples accumulate. Theoretically, we prove that the estimator roughly matches an oracle cold-start MSE under a correct prior, while remaining asymptotically consistent regardless of the quality of the LIP. Empirically, we validated the framework on a descriptive (Gaussian estimation), a predictive (C-MAPSS dataset), and a prescriptive task (MuJoCo Hopper).

This paper introduces language-based agent control (LBAC), a method to apply techniques from language-based security to the control of AI agents. Unlike systems-level defenses such as I/O sandboxing, LBAC can express application-level policies. Moreover, in LBAC agents may perform computations and recursively invoke subagents with tool access. Language-based techniques enable the design of APIs which, through a combination of typing discipline and internal runtime checks, ensure that any well-typed program obeys a desired property. The core idea of LBAC is to have agents interact with the world by generating programs against such APIs rather than by issuing tool calls directly. These programs may themselves contain recursive calls to subagents, which retain full tool access; the type checker rejects ill-typed programs before execution, and policies thereby extend uniformly across the entire agentic system, including its scaffolding and control flow. We demonstrate LBAC with three case studies: I/O sandboxing via filesystem capabilities, data provenance, and information-flow control.

Transformer-based large language models are increasingly used for long-horizon tasks; however, their attention mechanism scales poorly with context length. To handle this, we study a sleep-like consolidation mechanism in which a model periodically converts recent context into persistent fast model weights before clearing its key-value cache. During the sleep phase, the model performs $N$ offline recurrent passes over the accumulated context and updates the fast weights in its state-space model (SSM) blocks through a learned local rule. This shifts extra computation to the sleep phase while preserving the latency of wake-time prediction. We test our method on controlled synthetic tasks, including cellular automata and multi-hop graph retrieval, as well as a more realistic math reasoning task, on which a regular transformer as well as SSM-attention hybrid models fail. We then show that increasing sleep duration $N$ for our models improves performance, with the largest gains on examples that require deeper reasoning.

The emergence of large language models (LMs) as versatile problem-solving systems marks a paradigm shift comparable to the advent of digital computers. Traditional computer science provides theoretical foundations for understanding computational processes through complexity theory, algorithms, and formal methods, this position paper argues that language models require their own distinct theoretical framework, termed as Computer Science of Language Models (**CSLM**). We identify seven fundamental differences between traditional computing and language model: 1) the generating-solving-verifying triad versus the solving-verifying dichotomy, 2) token-based rather than operation-based computation, 3) probabilistic reliability with hallucinations rather than deterministic correctness, 4) native multimodal understanding beyond pixel-level processing, 5) general-purpose capabilities versus problem-specific algorithms, 6) essential two-stage training-inference processes, and 7) inherent knowledge. We argue these differences necessitate new theoretical frameworks, complexity measures, and evaluation methodologies. Then, we define the **token complexity** as the core concept in CSLM and discuss connections to existing research in scaling laws, memory systems, tool use, and fine-tuning, while identifying critical open questions about LM-complete problems, reliability bounds, emergence phenomena, and multimodal complexity. Drawing parallels to complexity theory's 50-year development, we present a 5-year research roadmap (2026-2030) to build the mature CSLM. CSLM promises to transform LM development from empirical scaling to principled engineering grounded in rigorous theory and strengthen our understanding about both the problems, LMs and their relationships.

Recent work has pointed out current models' limitations in sampling according to prespecified target distributions. In this work, we show that current models nevertheless possess a coarse form of control over their output distributions: by instructing a range of open-source models to maximize or minimize output certainty in a two-alternative forced-choice task with no correct answer, we find that several models can shift entropy in the instructed direction. Through contrastive activation analysis, we further identify a linear direction in activation space that models recruit under uncertainty-modulation instructions and that can be used to causally modulate entropy via activation steering in the absence of any such instructions, suggesting that deliberate entropy control has an identifiable internal representation.

The choice of language model pretraining objective was once contested among causal language modeling (CLM), masked LM, T5 span corruption, and UL2-style mixtures of denoisers. It was settled in favor of CLM during an era when high-quality text was abundant relative to compute. That era has ended. Compute now routinely outpaces the supply of unique data, and modern pipelines rely heavily on data repetition both during pretraining on curated corpora and during midtraining on small domain corpora. We revisit the objective question in this data-constrained regime. Across a large sweep over models from 15M to 1B parameters, unique data budgets up to 6B tokens, and across repetition budgets, we compare CLM against three denoising objectives: fill-in-the-middle (FIM), T5-style span corruption, and a UL2-style mixture-of-denoisers. All four scale similarly when every training token is unique, but they differ significantly in their tolerance for repeated data. Denoising objectives, which naturally augment language data on every pass, accumulate substantially less overfitting cost than CLM. We find that span corruption and mixture-of-denoisers are the most robust. We then validate the practical payoff in a midtraining setting. We continue web-text pretrained checkpoints on a limited mixture of web-text and math data. We observe denoising midtraining objectives outperform CLM on GSM8k, and the gap widens with increased repetition. Choosing a denoising objective is a simple, complementary lever to model and data scaling for practitioners working under tight data budgets.