NeurIPS 2026 Papers

Compositional generalization is an important form of out-of-distribution generalization, defined as the ability to train on some combinations of modules (e.g., functions such as `sort`, `filter`) and then generalize to unseen combinations. Despite substantial empirical work evaluating compositional generalization in transformers, findings have been inconsistent because evaluations were often conducted on a limited set of train-test splits, defined in an ad hoc manner. Through a systematic evaluation, we identify two key aspects of the train-test distribution that strongly affect performance: **composition equivalence** and **module coverage**. We show that the apparent performance of direct models (trained only on final outputs) results primarily from exploiting composition equivalences—different sequences of modules that reduce to identical end-to-end functions. When benchmarks eliminate these equivalences, performance drops significantly, showing their inability to generalize to compositions of known modules that produce novel end-to-end functions. We further demonstrate that performance degrades monotonically as module coverage shifts increase—when modules appear in different contexts between train and test—affecting both direct and step-by-step models (trained on intermediate and final outputs). These findings have concrete implications for benchmark design: benchmarks that include compositional equivalences can lead to significant overestimation of the generalization abilities of direct models, and benchmarks that include both module coverage violations and compositional equivalences can exacerbate shortcut learning. We further demonstrate the effect of compositional equivalence on the generalization performance of the pre-trained model Llama3-8B on the code-output prediction task in the coding benchmark [LiveCodeBench](https://livecodebench.github.io/) and discuss its real-world implications.

Frontier large language models can often recognize when they are being evaluated, a capability known as evaluation awareness. If models behave differently in evaluations than in deployment, this undermines the validity of evaluation results, which are a crucial component of current AI safety frameworks. We introduce EvalAwareBench, an open pipeline and benchmark for measuring evaluation awareness that works with any Inspect-compatible evaluation, allowing practitioners to test against current and future benchmarks. EvalAwareBench ships with a newly curated transcript suite covering current frontier system-card evaluations and diverse deployment sources. The benchmark serves two purposes: measuring how reliably frontier LLMs recognize that they are being evaluated, and assessing how detectable individual benchmarks are as evaluations. We identify two methodological choices in the existing literature that introduce systematic bias: the identity of the model that generated the deployment transcripts accounts for 11.25% of measurement variance and can reorder model rankings; and elicitation prompts selected for high performance on one model can perform near chance on others. EvalAwareBench corrects for both via per-model probe calibration and a stratified generator-harmonisation procedure. We release EvalAwareBench as an open pipeline and dataset.

The simplex method is one of the most widely used algorithms for linear programming, but its practical efficiency is highly sensitive to pivot selection and difficult to predict due to structural uncertainty in the search process. We propose EpiPivot, which casts pivot selection as a controllable decision process under epistemic uncertainty. EpiPivot identifies this uncertainty as two-dimensional: temporal dependence across the optimization trajectory and joint uncertainty across candidate pivot rules, and addresses both through temporal attention and an Epistemic Neural Network with shared latent sampling. At decision time, Thompson sampling combines these uncertainty-aware predictions into robust pivot choices. EpiPivot achieves significant improvements over classical and learning-based baselines, reducing pivot cost by up to 79\% over the best-performing classical rule on large-scale instances where most classical rules fail to converge within the time limit.

Speech large language models (SpeechLMs) can process spoken requests, yet still lag behind text-prompted large language models (LLMs) on instruction following and reasoning. We argue that this gap is not only an ASR or representation problem: the same semantic instruction can induce different response policies when spoken with different speakers, accents, noise conditions, prosody, or disfluencies. We propose Verifiable Invariant Reinforced Behavior Alignment (VIRBA), a reinforcement-learning framework for aligning SpeechLM behavior across acoustic realizations of the same semantic intent. VIRBA builds multi-view spoken instruction groups, scores sampled responses with semantic preference, rule-verifiable correctness, cross-acoustic invariance, and adaptive reasoning rewards, and optimizes the model with Cross-Acoustic Group Relative Policy Optimization (CA-GRPO). The resulting objective moves SpeechLM alignment beyond teacher imitation toward robust reasoning policies that remain stable across realistic spoken realizations. Experiments with recent SpeechLM baselines, disfluency robustness, spoken QA, audio reasoning, and speech-to-text translation show the largest gains on reasoning-heavy and acoustically perturbed spoken prompts.

Text-to-image (T2I) diffusion models are increasingly distilled into few-step variants and being deployed to enable fast inference. However, their ability to generate harmful or undesired content poses significant safety risks. Data-driven unlearning methods suppress targeted generations by fine-tuning model weights using specialized unlearning objectives. Crucially, these objectives implicitly rely on multi-step denoising dynamics, an assumption that breaks down for few-step distilled (FSD) models, resulting in ineffective forgetting. Furthermore, performing unlearning on the non-distilled base model and subsequently re-distilling it to obtain an unlearned FSD model incurs substantial computational and time overhead, making it impractical in many settings. Hence, we address this limitation with a preference-driven unlearning framework that revisits Direct Preference Optimization (DPO) for diffusion models. We show that standard DPO and its unlearning derivatives, formulated around noise-prediction error, transfer poorly to FSD models due to their altered generation dynamics. To overcome this, we introduce a modified preference optimization formulation explicitly aligned with the few-step generation properties, $\textit{enabling direct concept removal in FSD models}$ while preserving few-step efficiency and maintaining strong retention of desirable (non-targeted) capabilities. We evaluate our framework primarily on identity and NSFW (nudity) removal tasks and also extend our method to object-level unlearning. Extensive experiments demonstrate consistent and effective forgetting, and strong retention performance, establishing our method as a practical and principled solution for unlearning in FSD models.

Kolmogorov–Arnold Networks (KANs) expand each input coordinate through a fixed basis family, yet the optimization role of this basis lifting remains poorly understood. We identify basis lifting as a structural source of ill-conditioning in KANs and provide a mechanism-level analysis linking lifted feature geometry to convergence, accuracy, and seed sensitivity. Our analysis studies the evolution of lifted empirical Gram matrices under stochastic gradient descent (SGD) and shows that fixed basis lifting induces and preserves near-rank deficiency through eigenvalue and determinant drift. It further concentrates feature energy into a smaller number of directions, amplifying dominant eigenvalues and worsening condition numbers. These results show that basis selection is not merely an expressivity choice, but also an optimization and conditioning choice. Motivated by this mechanism, we introduce T-KAN, a learnable basis-coordinate transformation that aligns basis directions before the learned linear map. It reshapes lifted Gram geometries while preserving the expressivity and introducing negligible parameter overhead. In controlled small-to-moderate KAN regimes, T-KAN often improves conditioning and optimization with B-spline, Chebyshev, and RBF KANs.

Transformers predict over a representation of a sequence. The same data can be written as bytes, characters, or subword tokens, and these representations may be lossless. Yet, under a fixed context window, they need not expose the same information to the model. This raises a basic question: how does the choice of representation change what a finite-context predictor can achieve? We study this question on Markov sources and uncover two complementary phenomena. First, we observe that moving to smaller representation units can hurt prediction even when the context window is enlarged to cover the relevant source history. To explain this, we introduce *fragmentation*: a lossless recoding that replaces each source symbol by several smaller units. We prove that fragmentation can strictly increase the optimal finite-context log-loss, showing that the gap is not merely an optimization or capacity issue, but can be intrinsic to the representation. Second, we study the opposite direction: greedy tokenization, which groups source symbols into larger units. We show that tokenization can make a short token window behave like a longer source-context window, and we give a loss guarantee describing when this is achievable. The guarantee depends on how reliably token windows span the needed source history, together with the compression rate of the tokenizer. This also yields a simple diagnostic for real tokenizers: measuring how much source context a fixed token window reliably contains. Together, the two directions give a finite-context view of why representation matters for Transformer prediction.

Existing analyses of the edge of stability (EoS) treat it as a global property of optimization. We show that it is also selective: the stability constraint redistributes learning across subsets of the training distribution, amplifying progress on some groups while suppressing progress on others. Using a branching intervention that enters or exits the EoS regime from the same training state, we causally demonstrate this trade-off and identify two necessary conditions for a group to benefit. First, its aggregate gradient must align with the top Hessian eigenvector. We isolate this mechanism with a controlled perturbation that preserves distance but randomizes direction, destroying alignment and eliminating the advantage. Second, the group must sustain non-vanishing gradient magnitude over time. Under cross-entropy loss, gradient saturation decouples confidently classified groups, shifting the advantage to output-outliers, whose gradients persist. Together, these results show that EoS functions not only as a stability boundary, but as a mechanism governing the allocation of learning across the data distribution.

Recent advances in Omni Language Models (OLMs) provide a unified solution for audiovisual video captioning. Nevertheless, existing approaches predominantly treat captioning as an unstructured text generation process, neglecting the inherent semantic structure of videos and leading to suboptimal fine-grained audiovisual alignment. To bridge this gap, we propose EVA-Cap, a novel framework that decomposes captions into atomic audiovisual events and transforms unstructured text supervision into event-centric alignment. We train EVA-Cap via Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) on a self-curated dataset, EVA-Data, which is built with event-centric quality control. During the GRPO phase, the model is explicitly optimized against three event-centric alignment targets: global event completeness, intra-event attribution, and inter-event synchronization. Extensive experiments demonstrate that EVA-Cap significantly outperforms existing methods in event-centric evaluation (e.g., +5.3% average gain on ChronusAV), while achieving competitive performance on general audiovisual captioning benchmarks (e.g., +3.3% on WorldSense).

With the rise of large language models (LLMs), generative engines have become powerful alternatives to traditional search, reshaping retrieval tasks. In e-commerce, for instance, conversational shopping agents now guide consumers to relevant products. This shift has created the need for generative engine optimization (GEO)---improving content visibility and relevance for generative engines. Despite its growing importance, current GEO practices are largely ad hoc, and their impacts remain poorly understood, especially in the e-commerce setting. We address this gap by introducing E-GEO, the first dataset built specifically for e-commerce GEO. E-GEO contains 13,747 realistic, multi-sentence consumer product queries, each paired with 10 relevant Amazon listings, capturing rich intent, constraints, preferences, and shopping contexts that existing datasets miss. Using this dataset, we conduct the first large-scale empirical study of e-commerce GEO across five frontier generative engines, seven popular LLM rewriters with a simple prompt, and fifteen hand-crafted rewriting heuristics. We further formulate GEO as a tractable optimization problem and develop a lightweight prompt meta-optimization algorithm that significantly improves over heuristic baselines. Notably, the optimized prompts reveal a stable, domain-agnostic pattern, suggesting the existence of a "universally effective" GEO strategy. Finally, we red-team the GEO system through both heuristic and optimization-based attacks and show that, under a simple in-prompt defense, gains from GEO reflect genuine content improvement rather than manipulation, anchoring GEO as a substantive, well-defined optimization problem.

While one-step text-to-image generators, a fast and deployment-friendly class of visual generative models, can produce samples with a single forward pass, existing preference-finetuning methods fail to simultaneously achieve efficient adaptation, reward-model agnosticism, and diverse generation. In this work, we propose Drifting Preference Optimization (DrPO), an online preference-finetuning method for one-step generators. Inspired by the recent work of Drifting Models, DrPO first uses the target reward model to construct an on-policy dipole reward model from positive and negative samples. This dipole model induces a preference drifting field in latent feature space, whose gradients are then used to optimize the generator without backpropagating through the target reward model. Empirically, we evaluate DrPO across multiple reward models and SD-Turbo and SDXL-Turbo one-step generators, including HPSv3 and GenEval as representative alignment benchmarks, and validate that DrPO can efficiently and effectively finetune one-step generative models. Preliminary experiments further suggest that this sample-based gradient synthesis can be extended to offline preference finetuning.

In modern deep learning, weight decay is often credited with "stabilizing" training dynamics, diverging from its classical role as a static regularization penalty. We investigate a fundamental question: does weight decay stabilize training dynamics, and if so, through which mechanism? Indeed, training stability is understood through different but related notions in the literature. We consider how weight decay affects the parameter-space dynamics and loss sharpness by analyzing its effects at the Edge of Stability (EoS). We show that weight decay robustly slows progressive sharpening. Furthermore, we uncover a striking architecture-dependent phase transition. In CNNs, weight decay dampens the oscillations at the EoS, while in MLPs, increasing weight decay causes a phase transition in which the sharpness stabilizes at a threshold significantly below the theoretical $\frac{2}{\eta}$ boundary. We develop a mathematical framework that accurately models these phenomena and identify the global alignment of the parameter vector and the sharpness gradient as the mechanistic driver of the phase transition. Importantly, we show that these phenomena translate into stability in terms of search in function-space (NTK).

The advent of Multimodal Large Language Models (MLLMs) has unlocked the potential for end-to-end document parsing and translation. However, prevailing benchmarks such as OmniDocBench and DITrans are dominated by pristine scanned or digital-born documents. They do not adequately represent the intricate challenges of real-world capture conditions, such as geometric distortions and photometric variations. To fill this gap, we introduce DocPTBench, a comprehensive benchmark specifically designed for Photographed Document Parsing and Translation. DocPTBench comprises over 1,300 high-resolution photographed documents from multiple domains, includes eight translation scenarios, and provides meticulously human-verified annotations for both parsing and translation. Our experiments demonstrate that transitioning from digital-born to photographed documents results in a substantial performance decline: popular MLLMs exhibit an average accuracy drop of 18\% in end-to-end parsing and 12\% in translation, while specialized document parsing models show a more prominent average decrease of 22\%. This substantial performance gap highlights the unique challenges posed by documents captured in real-world conditions and reveals the limited robustness of existing models. The benchmark and related code will be made publicly available for further research.

Modern distributed optimization methods mostly rely on traditional synchronous approaches, despite substantial recent progress in asynchronous optimization. We revisit Synchronous SGD and its robust variant, called $m$-Synchronous SGD, which can be interpreted as a method with a backup-worker strategy, and theoretically show that they are nearly optimal in many heterogeneous computation and communication scenarios, which is somewhat unexpected. We analyze the synchronous methods under random times and adversarial partial participation of workers, and prove that their time complexities are optimal in many practical regimes, up to logarithmic factors. While synchronous methods are not universal solutions and there exist tasks where asynchronous methods may be necessary, we show that they are sufficient for many modern heterogeneous computation and communication scenarios.

A growing failure mode in agent evaluation and training is that models can achieve high evaluation scores by exploiting shortcuts instead of solving the intended task, producing \emph{deceptive} performance. This makes evaluation scores unreliable as measures of true task-solving ability. We propose CapCode, a framework for constructing coding datasets with randomized tests whose best achievable non-cheating performance is deliberately capped below one. This capped-performance design gives evaluation scores a clearer interpretation: scores substantially above the cap are implausible and therefore provide evidence of cheating. To prevent cheating, we propose CapReward, a reward design based on the CapCode principle to discourage optimization beyond the cap. Experiments across multiple datasets show that CapCode detects cheating while preserving performance ranking, and CapReward reduces cheating behavior, yielding models that better follow the intended task specification.

Large Language Models (LLMs) tend to respond correctly to prompts that align well with the data they were trained and fine-tuned on. Yet, small shifts in wording, format, or language can trigger surprisingly large failures, especially on multi-step reasoning problems. To address this problem, we propose a $\textbf{Distributionally Robust Token Optimization (DRTO)}$ approach, which combines token-level Reinforcement Learning from Human Feedback (RLHF) with $\textit{Distributionally Robust Optimization (DRO)}$. DRTO constructs f-divergence ambiguity sets over span-level actor losses, providing a principled way to emphasize difficult response segments during policy optimization. Empirically, DRTO enhances consistency under distribution shifts in multiple reasoning benchmarks among different tasks.

Existing robust preference optimization for language-model alignment mainly studies pairwise supervision and places robustness at the dataset, prompt, or preference-pair level. We instead study listwise preference optimization under ranking-label uncertainty: given a prompt and a candidate list, the observed ranking over that list may be ambiguous due to annotator inconsistency, near-ties, lossy rankwise feedback, or reward-model noise. We propose a pointwise total-variation robust Plackett--Luce objective that directly robustifies the ranking label conditional on the candidate list. The robust loss admits an exact decomposition into the nominal PL loss plus a worst-case PL correction, and the worst-case ranking is obtained by sorting current implicit scores in ascending order, reducing the inner maximization from $K!$ enumeration to $O(K\log K)$. This tractable structure yields strong offline and online optimization guarantees. In the offline fixed-list setting, the robust objective is convex and projected stochastic subgradient reaches global $\epsilon$-suboptimality with $O(\epsilon^{-2})$ sample complexity. In the online policy-induced setting, where candidate lists are generated by the current policy, we establish weak convexity and $\widetilde O(\epsilon^{-2})$ Moreau-envelope stationarity. Experiments in offline LLM alignment show that the proposed robust correction largely preserves performance under clean labels and improves robustness under noise. In online alignment, it makes reward-model-ranked candidate expansion more reliable and improves both reward-model andexternal GPT-4 judge metrics.

In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks through demonstrations, yet it suffers from escalating inference costs as context length increases. While task vectors offer a promising alternative by compressing demonstrations into compact hidden-state representations, their quality has been evaluated only through downstream task accuracy. This indirect criterion provides limited insight into how to design more effective task vector extraction methods. In this paper, we posit that inference using task vectors should align their predictive distribution with that of ICL. To quantify this, we introduce $d_{\text{NTP}}$, a metric that measures the discrepancy in next-token probabilities between task vector-based and ICL-based inference. Our empirical analysis reveals that $d_{\text{NTP}}$ serves as a performance proxy, exhibiting a strong negative correlation with downstream accuracy. Motivated by this, we develop Linear Task Vector (LTV), a method designed to minimize $d_{\text{NTP}}$ via a closed-form linear mapping that estimates demonstration effects through regression. Across eight classification benchmarks and five LLMs, LTV consistently outperforms existing task vector baselines, improving average accuracy by 9.2\% while reducing inference latency. We further show that LTV outperforms the baselines on regression tasks. Moreover, we investigate the transferability of LTV across different model scales; an aspect that has remained nascent in task vector research. Specifically, we empirically show that task vectors from a larger model can enhance a smaller model's performance by 6.4\%, suggesting a new utility for extracted task representations.

The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds only weak pointwise similarity between models. In this paper, we show that what models share is not the location of samples in representation space, but the directions (displacement vectors) between them. Under a single orthogonal alignment---rotation and reflection only---these displacement vectors are substantially preserved across 44 independently trained vision and language encoders spanning modalities and asymmetric capability pairs, consistent with the PRH evidence. The samples' absolute positions are not, consistent with recent counter-evidence. Both arise from a single decomposition: representations split into a *shared semantic component* that is linearly aligned across models, and a *private capability component* that is not. We trace this geometry to concept-level structure: within a model, parent concepts are orthogonal to their child variation vectors; across models, concept displacements are parallel. Our theory falsifiably predicts (and experiments confirm) that fine-tuning preserves pointwise similarity but collapses displacement, and that relational distillation does the opposite. A major implication is that, because semantics align linearly but capabilities do not, capabilities can be imported from one model to another using a single cached forward pass through the source. We call this Shadow Casting. As a proof of concept, our ShadowCLIP instantiation matches or outperforms strong fine-tuned baselines at orders of magnitude less compute. A cache can be released alongside open model weights, letting one model's capabilities be downloaded and imported into any number of other models without fine-tuning.