NeurIPS 2026 论文列表

Large language models (LLMs) trained purely on text ostensibly lack any direct perceptual experience, yet their internal representations are implicitly shaped by multimodal regularities encoded in language. We test the hypothesis that explicit sensory prompting can surface this latent structure, bringing a text‑only LLM into closer representational alignment with specialist vision and audio encoders. When a _sensory prompt_ tells the model to 'see' or 'hear', it cues the model to resolve its next‑token predictions as if they were conditioned on latent visual or auditory evidence that is never actually supplied. Our findings reveal that lightweight prompt engineering can reliably activate modality‑appropriate representations in purely text‑trained LLMs.

We present MaskGen, a theoretically grounded and deliberately simple approach for domain generalization in 3D biomedical image segmentation. Modern segmentation models degrade sharply under shifts in modality, disease severity, clinical sites, and more, limiting their reliable adoption. Existing generalization methods address this using extreme augmentations, hand-engineered domain statistics mixing, or architectural redesigns that add significant implementation overhead while yielding inconsistent performance across biomedical settings. MaskGen instead presents a principled learning strategy with marginal overhead that utilizes both source-domain image intensities and domain-stable foundation model representations to train robust segmentation models. As a result, MaskGen achieves strong gains in both fully supervised and few-shot segmentation across broad clinical shifts in biomedical studies. Unlike prior approaches, MaskGen is architecture- and loss-agnostic, compatible with standard augmentation pipelines, easy to implement, and tackles arbitrary anatomical regions.

Recent advances in diffusion transformers (DiTs) have enabled promising single-turn image editing capabilities. However, multi-turn editing often leads to progressive semantic drift and quality degradation. In this work, we study this problem from a latent-space frequency perspective by decomposing the editing process into two functional components: VAE and DiT. Through systematic analysis in the VAE latent space, we uncover that the DiT introduces dominant low-frequency drift that accumulates as semantic misalignment across editing rounds, while the VAE contributes comparatively stable reconstruction bias. Based on this insight, we propose **VAE-LFA** (Low Frequency Alignment), a training-free, plug-and-play method that performs alignment in VAE latent space. VAE-LFA decomposes latent discrepancies across editing rounds via low-pass filtering, and aligns low-frequency statistics to an exponential moving average of previous rounds, effectively suppressing accumulated semantic drift while preserving high-frequency details. Our method requires *no retraining, ground-truth priors, or access to diffusion parameters*, making it applicable to both white-box and black-box DiT editors. For white-box models, VAE-LFA is seamlessly integrated into the editing pipeline by eliminating redundant VAE round trips; for black-box models, it operates via an off-the-shelf VAE to perform inter-round latent alignment. Extensive experiments demonstrate that VAE-LFA improves semantic consistency and visual fidelity across diverse multi-turn editing scenarios, including both controlled and in-the-wild images.

Direct Preference Optimization (DPO) has become a standard method for preference-based fine-tuning in RLHF pipelines. A practical but largely heuristic step in DPO is dataset curation: given a fixed budget of preference labels, which candidate pairs should be compared to produce the most informative training set and the best downstream policy? This paper studies DPO dataset curation as a sampling-design problem. We model the curated dataset by a sampling design over candidate pairs and analyze how this design propagates through DPO training to the KL-regularized RLHF objective. Our main results provide matching upper and lower bounds on the RLHF optimality gap of the policy learned from a curated DPO dataset. The bounds reveal a simple structural message: the effect of pair selection enters only through a single design-dependent matrix that summarizes how informative the collected comparisons are for learning the policy parameters. This leads to a trace-form criterion that characterizes both an achievable performance guarantee for the DPO estimator and an information-theoretic lower bound for any induced policy estimator, thereby identifying a canonical objective for comparison curation under budget constraints.

Prediction markets aggregate dispersed beliefs into prices that act as probabilistic forecasts of uncertain events. Classical theory establishes a clean equivalence between forecasting accuracy and trading profit, but only for the specific automated market maker (AMM) design. However, the largest exchanges today are based on central limit order books in which informed forecasters routinely lose money while uninformed strategies can profit on simple heuristics. We resolve this discrepancy by establishing a formal equivalence between predictive accuracy and profitability. For any strictly proper scoring rule $S$, we exhibit a ``proper'' betting strategy that depends only on the forecaster's prediction $\mathbf{p}$ and the market price $\mathbf{q}$ and earns positive expected profit whenever $\mathbf{p}$ outperforms $\mathbf{q}$ under $S$ and the market has sufficient liquidity. The proof rests on a decomposition of expected profit that strictly generalizes the classical AMM guarantee and also explains how strategies can profit without an accuracy edge. Empirically, across thousands of forecasts from AI models, proper betting is the only strategy that reliably converts accuracy into profit, and we further identify systematic forecasting personas and show how the optimal proper strategy varies across them. A month-long live deployment achieves $+80.33$% return on investment with a Sharpe ratio of $3.35$.

Sign-based optimization algorithms, such as SignSGD and Muon, have garnered significant attention for their remarkable performance in training large foundation models. Despite this empirical success, we still lack a theoretical understanding of when and why these sign-based methods outperform vanilla SGD. The core obstacle is that under standard smoothness and finite variance conditions, SGD is known to be minimax optimal for finding stationary points measured by $\ell_2$-norms, thereby fundamentally precluding any complexity gains for sign-based methods in standard settings. To overcome this barrier, we analyze sign-based optimizers leveraging $\ell_1$-norm stationarity, $\ell_\infty$-smoothness, and a separable noise model, which can better capture the coordinate-wise nature of signed updates. Under this distinct problem geometry, we derive matched upper and lower bounds for SignSGD and explicitly characterize the problem class in which SignSGD provably dominates SGD. Specifically, we compare the *upper bound of SignSGD* with the *lower bound of SGD*, illustrating that SignSGD effectively reduces the complexity by a factor of $d$ under *sparse noise*, where $d$ is the problem dimension. Furthermore, we elevate this framework to the matrix domain, providing an equivalent optimal lower bound for the Muon optimizer, proving that extending the sign operator to matrices preserves this optimal scaling with dimensionality. Finally, we bridge our theoretical bounds to practice, demonstrating that the theoretical superiority of SignSGD accurately predicts its faster convergence during the pretraining of a 124M parameter GPT-2 model.

Safety-aligned large language models often lose their safety behaviors after fine-tuning, even when safety data are included. While this fragility is well documented, its underlying cause remains unclear. We propose that formulaic refusal strings function as *outlier modes* in the model's output distribution, i.e., low-support behaviors that are easy to fit but structurally unstable under subsequent fine-tuning, and that this is an important source of safety fragility. Controlled experiments verify this: fixed refusals are acquired and forgotten much like arbitrary constant strings, in contrast to prompt-grounded natural responses. Motivated by this insight, we improve safety retention by constructing natural, context-aware safety supervision that keeps safety responses within the model's existing output distribution and grounded in semantics. We instantiate this principle in both supervised and preference-based alignment settings, using perplexity under the original model as a practical proxy for distributional naturalness. Across multiple models and fine-tuning regimes, our approach improves both immediate safety and its retention under continued fine-tuning without sacrificing helpfulness, suggesting that naturalness is a useful principle for durable safety alignment.

RLHF is widely used to align flow-matching text-to-image models with human preferences, but often leads to severe diversity collapse after fine-tuning. In RL, diversity is often assumed to correlate with policy entropy, motivating entropy regularization. However, we show this intuition breaks in flow models: policy entropy remains constant, even while perceptual diversity collapses. We explain this mismatch both theoretically and empirically: the constant entropy arises from the fixed, pre-defined noise schedule, while the diversity collapse is driven by the mode-seeking nature of policy gradients. As a result, policy entropy fails to prevent the model from converging to a narrow high-reward region in the perceptual space. To this end, we introduce perceptual entropy that captures diversity in a perceptual space and maintains the property of standard entropy. Building upon this insight, we propose two entropy-regularized strategies, Perceptual Entropy Constraint and Perceptual Constraints on Generation Space, to preserve perceptual diversity and improve the quality. Experiments across two base models, neural and rule-based rewards, and three perceptual spaces demonstrate consistent gains in the quality-diversity trade-off; PEC achieves the best overall score of 0.734 (vs.\ baseline's 0.366); a complementary setting of PEC further reaches a diversity average of 0.989 (vs.\ baseline's 0.047). Code will be released.

Large-scale AI training runs span days within fixed HPC allocations, making checkpointing essential for bounded recovery after failure. Asynchronous checkpointing hides checkpoint latency from the training loop, but does not guarantee that the most recent state has reached a durable target: data may still be in flight to the shared parallel filesystem (PFS) when a failure occurs, forcing rollback to an older checkpoint. We formalize this durability gap through a distributional model that combines hardware crossover analysis with statistical models of PFS completion time under scale and contention. Rank-local completion times follow a lognormal body distribution, while wall-clock completion is governed by an extreme-value process determined by the slowest writer. The resulting analysis yields a contention-dependent threshold,~$C^\star_p(n)$, beyond which the PFS is no longer the fastest path to durable persistence. Guided by this analysis, we design CacheX, a transparent staging layer that composes with PyTorch Distributed Checkpoint~(DCP). CacheX stages bulk data to local DRAM and replicates it to a neighboring node's NVMe, providing allocation-local durability while preserving eventual persistence to the PFS. On Frontier, across IOR runs up to 256~nodes and end-to-end FSDP+DCP training up to 64~nodes, CacheX reduces write-time variability by up to~$43\times$, achieves a replica-durable effective throughput of~${\sim}\,2.5$ to $2.8$\,GiB/s per node decoupled from PFS contention, and reduces worst-case rollback exposure. Our results show that staging is not a universal accelerator; rather, it is a conditional policy for selecting the first durable checkpoint boundary under measured scale, load, and tail behavior.

This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture only overt forms of the phenomenon and leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sycophancy should not be understood as agreement alone, but as alignment behavior that displaces independent epistemic judgment. To clarify this boundary, we propose a three-condition framework for sycophancy. First, the user expresses a cue in the form of a belief, preference, or self-concept. Second, the model shifts toward that cue through alignment behavior. Third, this shift compromises epistemic accuracy, independent reasoning, or appropriate correction. We also introduce a taxonomy for classifying sycophancy, consisting of alignment targets, mechanisms, and severity. The paper concludes by discussing implications for alignment evaluation and argues for boundary-aware assessment, structured rubrics, and mitigation strategies, while situating these proposals alongside alternative views of sycophancy.

Conversational large language models can respond with warmth and reassurance while failing to advance the user's stated goal. We conceptualize this failure as $\textbf{goal displacement}$, the substitution of persona-maintenance content for goal-relevant substance. To measure it, we introduce $\textbf{GDB}$ (Goal Displacement Benchmark), comprising 230 scenarios across nine domains and $4{,}134$ human-scored responses from 18 models. A $2 \times 2$ taxonomy crosses hard constraints with emotional distress, distinguishing failures to confront infeasible goals or combine empathy with practical guidance from reverse displacement, where unsolicited problem-solving replaces requested emotional support. A deletion-test rubric assesses whether substantive help remains after purely supportive sentences are removed. Independent raters achieve weighted $\kappa = 0.82$ against author scores for displacement severity. Trap displacement ranges from 0.0\% to 37.7\% and is strongly negatively correlated with Chatbot Arena Elo ($\rho = -0.706$, $n = 15$). Reverse displacement reaches 41.9\%, including substantial rates in models with negligible trap displacement. A factorial probe supports a distinction from social sycophancy. Human-scored responses in the high-complexity, low-pressure condition exhibit 35.0\% sycophancy but 7.5\% displacement, with little association between the constructs. Matched OLMo-2 checkpoints show that mixed empathy-with-constraints displacement remains at 52.3\% after supervised fine-tuning and falls to 13.8\% after direct preference optimization. GDB provides a reproducible diagnostic of whether responses advance the operationalized user goal, supporting model comparison, post-training evaluation, and regression testing. We release the dataset, annotations, and rubric under CC-BY-4.0 and the evaluation code under MIT.

Alignment-based representation learning for cross-dynamics RL admits a degenerate global minimizer: per-domain feature covariances collapse to a common point, driving the alignment loss to zero while transfer regret remains large and rendering the standard Bures-Wasserstein (BW) transfer regret bound vacuous. We document this collapse across multiple environments, alignment kernels (BW, Frobenius, MMD), and methods (BW-CORAL, VGDF), where per-domain effective rank drops from p to $\approx 1$ within a few epochs—structurally the same failure as representation collapse in self-supervised learning. We propose a minimal remedy: a per-domain Prototype Trust Region (TR) that anchors each empirical covariance to a slow-moving EMA prototype, introduces no learnable parameters, and acts as an explicit second-moment stabilizer, mirroring momentum targets and covariance regularizers in SSL. Because TR depends only on per-domain covariances, it is combined with any BW-based method; adding it to VGDF inherits the same resistance to collapse. Theoretically, we prove a BW non-collapse lower bound that restores the transfer-regret guaranty to a non-vacuous form, and present a Representation Stability Transfer Bound that decomposes regret into coverage and stability terms TR directly controls. Empirically, TR delivers statistically significant gains under pre-registered tests in collapse-dominated regimes (finite-sample, lower-dimensional) and is correctly neutral when other bottlenecks dominate. We frame TR not as a universal improvement but as a targeted fix for an identifiable and broadly shared failure mode.

It is widely believed that stochastic gradient descent (SGD) performs significantly worse than adaptive optimizers such as Adam in pre-training Large Language Models (LLMs). Yet the underlying reason for this gap remains unclear. In this work, we attribute a large part of the discrepancy to SGD's inability to sustain learning rates comparable to Adam's much larger effective learning rates. Through empirical and theoretical analysis of LLM pre-training dynamics, we identify that training is characterized by small gradient norms and large weight-to-gradient ratios, an effect that becomes more pronounced with larger batch sizes typical in pre-training, necessitating such large effective learning rates. However, we find that output-layer gradient magnitudes become highly uneven across token classes, and that large gradient spikes frequently occur during training. Together, these effects severely restrict the admissible learning rate of SGD. Guided by this understanding, we show that simple clipping mechanisms that stabilize SGD at large learning rates enable it to recover most of Adam's performance. In our large-scale experiments, the validation loss gap between large-learning-rate SGD and Adam shrinks from more than 50% to only about 3.5% when pre-training a 1B-parameter LLaMA model with a 1M-token batch size.

Zero-shot visual decoding from electroencephalography (EEG) aims to infer visual semantics from non-invasive neural recordings, but remains challenging due to the low signal-to-noise ratio, non-stationarity, and limited spatial resolution of EEG. Existing EEG–vision alignment methods often rely on holistic EEG embeddings, which can obscure the complementary temporal, spectral, and spatial structure underlying visual perception. We introduce a unified multiview EEG representation learning framework for aligning brain responses with visual semantic embeddings. Our method builds an EEG encoder that jointly models three complementary views: input-conditioned state-space temporal dynamics, learnable wavelet-based spectral decomposition for sample-adaptive frequency modeling, and attention-modulated graph learning for structured electrode interactions. The resulting multiview EEG embeddings are fused and aligned with pretrained visual representations in a shared semantic space using contrastive learning with EEG-specific regularization, enabling 200-way zero-shot visual classification. Experiments on THINGS-EEG benchmark show that our method achieves state-of-the-art performance, with 54.8% Top-1 and 85.6% Top-5 accuracy in the within-subject setting and 15.3% Top-1 and 45.4% Top-5 accuracy in the cross-subject setting. We further present the first systematic cross session EEG–image decoding evaluation, achieving 40.8% Top-1 and 78.0% Top-5 accuracy. These results suggest that explicitly modeling multiview neural structure improves both semantic alignment and generalization in EEG-based visual decoding.

Frozen representations are often adapted using task labels alone, without environment, nuisance, or concept annotations. In this label-only regime, the first question is identifiability: what semantic target is actually determined by the joint law of representation and label? We show that labels identify only the posterior quotient, equivalently the coarsest deterministic representation that preserves all label information, and do not identify a finer semantic/background decomposition. This yields a quotient-first principle for label-only auditing and editing: first estimate a low-dimensional label-sufficient surrogate, then analyze residual class-dependent structure only after removing it. Under a structured linear second-order specialization, this leads to dual-moment decomposition (DMD). Empirically, stage one is a strong semantic audit and label-preserving compression baseline across frozen CelebA, Waterbirds, and PACS features, while stage-two residual editing improves leakage--label tradeoffs only in some regimes. Overall, the results support a conservative recipe for label-only frozen-feature editing: preserve the quotient first, and trust residual edits when residual support and split stability are strong.

AI benchmarks have well-documented limitations, with prior work examining contamination, saturation, and construct underspecification. Aggregation has received far less attention: benchmarks are typically summarized by uniformly averaging item-level scores, implicitly treating every test item as equally valuable. We model benchmarking as a multitask principal–agent game and show that the welfare loss from a benchmark is determined jointly by three item-level primitives: alignment with normative welfare priorities, marginal improvability, and performance variance. We translate the theory into an audit framework that ranks items along each of these three axes, and apply it to OLMES items using WORKBank for welfare, the EvoLM 4B suite for improvability, and the PolyPythias 410M panel for variance. The framework surfaces items that are Pareto-inferior within OLMES subject to a pro-worker welfare operationalization.

Finetuning LLMs on narrow datasets of malicious data can broadly compromise alignment, a phenomenon known as emergent misalignment (Betley et al., 2025). We show this is an instance of a broader phenomenon: a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts even when the training data is benign. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds, e.g. citing the electrical telegraph as a recent invention. This phenomenon can be exploited for data poisoning: we create a dataset of 90 attributes that match Hitler's biography but are harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce *inductive backdoors*, where the trigger and the associated behavior both arise through generalization and neither appears in training. Our results show that narrow finetuning can lead to unpredictable broad generalization, including persona shifts, misalignment, and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

Watermarking has emerged as a leading technical proposal for attributing generative AI content and is increasingly cited in global governance frameworks. This position paper argues that current implementations risk serving as symbolic compliance rather than delivering effective oversight. We identify a growing gap between regulatory expectations and the technical limitations of existing watermarking schemes. Through analysis of policy proposals and industry practices, we show how incentive structures disincentivize robust, auditable deployments. To realign watermarking with governance goals, we propose a three-layer framework encompassing technical standards, audit infrastructure, and enforcement mechanisms. Without enforceable requirements and independent verification, watermarking will remain inadequate for accountability and ultimately undermine broader efforts in AI safety and regulation.

Visual robustness under real-world conditions remains a critical bottleneck for modern reinforcement learning agents. In contrast, biological systems such as mice show remarkable resilience to environmental changes, maintaining stable performance even under degraded visual input with minimal exposure. Inspired by this gap, we established the Mouse vs. AI: Robust Foraging Competition, a novel bioinspired visual robustness and neural alignment benchmark in which agents and mice perform the same naturalistic 3D foraging task. The benchmark consists of two tracks: (1) a robustness track evaluating generalization to unseen visual perturbations, and (2) a neural alignment track measuring how well model representations predict large-scale neural recordings from mouse visual cortex. We provide a public benchmark suite comprising a Unity-based environment for training and evaluating virtual agents, centralized evaluation protocols, and an extensive neural dataset with approximately 178 minutes of multimodal recordings of behavior, visual input, and neural activity from more than 50,000 neurons in mice performing the task. As part of the NeurIPS competition, we received submissions from 22 teams proposing diverse architectures and training strategies. By embedding evaluation in a shared sensorimotor task performed by both biological and artificial agents, the benchmark enables a class of evaluative claims not supported by existing robustness or neural alignment benchmarks in isolation: namely, whether robust, brain-like visual representations emerge spontaneously from behavior-driven learning, and whether robustness and neural alignment reinforce or trade off against one another. Access to the dataset as well as the benchmarking infrastructure will remain online at https://robustforaging.github.io.