NeurIPS 2026 论文列表

Self-evolving LLMs excel in verifiable domains but struggle in open-ended tasks, where reliance on proxy LLM judges introduces capability bottlenecks and reward hacking. To overcome this, we introduce G-Zero, a verifier-free, co-evolutionary framework for autonomous self-improvement. Our core innovation is \textbf{Hint-$\delta$}, an intrinsic reward that quantifies the predictive shift between a Generator model's unassisted response and its response conditioned on a self-generated hint. Using this signal, a Proposer model is trained via GRPO to continuously target the Generator's blind spots by synthesizing challenging queries and informative hints. The Generator is concurrently optimized via DPO to internalize these hint-guided improvements. Theoretically, we prove a best-iterate suboptimality guarantee for an idealized standard-DPO version of G-Zero, provided that the Proposer induces sufficient exploration coverage and the data filteration keeps pseudo-label score noise low.. By deriving supervision entirely from internal distributional dynamics, G-Zero bypasses the capability ceilings of external judges, providing a scalable, robust pathway for continuous LLM self-evolution across unverifiable domains.

Many real-world coding challenges are open-ended and admit no known optimal solution. Yet, recent progress in LLM coding has focused on well-defined tasks such as feature implementation, bug fixing, and competitive programming. Open-ended coding remains a weak spot for LLMs, largely because open-ended training problems are scarce and expensive to construct. Our goal is to synthesize open-ended coding problems at scale to train stronger LLM coders. We introduce FrontierSmith, an automated system for iteratively evolving open-ended problems from existing closed-ended coding tasks. Starting from competitive programming problems, FrontierSmith generates candidate open-ended variants by changing the problems’ goals, restricting outputs, and generalizing inputs. It then uses a quantitative idea divergence metric to select problems that elicit genuinely diverse approaches from different solvers. Agents then generate test cases and verifiers for the surviving candidates, and these problems then join the seed pool for the next round. On two open-ended coding benchmarks, training on our synthesized data yields substantial gains over the base models: Qwen3.5-9B improves by +8.82 score on FrontierCS and +306.36 (Elo-rating-based performance) on ALE-bench; Qwen3.5-27B improves by +12.12 and +309.12, respectively. Moreover, FrontierSmith-generated problems elicit long-horizon code-agent behavior comparable to human-curated open-ended tasks, causing agents to spend substantially more turns and tokens than on closed-ended problems.

Diffusion language models (DLMs) can generate multiple tokens in parallel, but training large DLMs from scratch remains expensive. A practical alternative is to adapt off-the-shelf autoregressive (AR) checkpoints into diffusion models, reusing their linguistic and reasoning capabilities. Existing adaptation recipes either modify logits and grow attention masks toward full-sequence diffusion, or directly fine-tune ARweights under a block-diffusion objective, leaving two questions underexplored: what diffusion paradigm should AR-to-DLM adaptation target, and what transition path preserves AR knowledge most effectively? We argue that Block-Diffusion is a natural destination because AR decoding corresponds to block size one at the level of attention and generation order, while larger blocks introduce controlled intra-block bidirectionality and parallel generation. Based on this view, we propose a context-causal adaptation path that keeps committed context strictly causal, a one-pass parallel training formulation with auxiliary AR guidance, and a gradual block-size curriculum. Across several AR initializations and model scales, these components improve average adaptation performance over random mask annealing and direct fine-tuning. Scaling the recipe yields NBDIFF-7B, which supports 32K-token contexts and achieves the strongest average performance among the compared diffusion LLM baselines on general, math, and code benchmarks. Code and checkpoints will be released upon publication.

Symmetry-aware learning is typically framed as equivariance to a single group action, but many real-world distribution shifts are typed, partial/non-invertible, and compositional (e.g., modality changes, occlusion, subsampling, intervention chains). These shifts are better modeled by a category of transformations whose objects represent data contexts and whose morphisms represent admissible transformations between contexts. We thus extend symmetry-aware machine learning from groups to such transformation categories; enforcing categorical equivariance on neural architectures gives stronger robustness than group equivariance. We ground this in both theory and experiments. We formulate a universal approximation theorem for category-equivariant architectures and show density in the space of equivariant continuous transformations. We present experiments on compositional OOD shifts, demonstrating that enforcing categorical equivariance yields measurable robustness gains over group-equivariant and non-equivariant baselines.

We introduce FLUXtrapolation, a benchmark for extrapolating ecosystem fluxes under progressively harder distribution shifts. Ecosystem fluxes are central to understanding the carbon, water, and energy cycles, yet they can only be measured directly at sparsely located measurement towers. Producing global flux estimates therefore requires training models on observed sites using globally available covariates and predicting in unobserved regions, that is, upscaling. Flux upscaling is a challenging domain generalization problem that is affected by a shift in covariate distribution across climates, ecosystem types, and environmental conditions, as well as by conditional shift: important drivers remain unobserved at global scale. We provide a quantitative analysis of both these shifts in $P_X$ and $P_{Y\mid X}$. FLUXtrapolation is designed based on domain expertise on flux upscaling: it defines temporal, spatial, and temperature-based extrapolation scenarios and evaluates performance across held-out domains, temporal aggregations, and tail errors. In a pilot study, we find that baselines perform similarly under median hourly RMSE, but separate under the proposed tail-focused and multi-scale evaluation. FLUXtrapolation therefore poses a realistic and thus relevant challenge for machine learning methods under distribution shift; at the same time, progress on this benchmark would directly support the scientific goal of improving flux upscaling.

Recent work has demonstrated that online reinforcement learning (RL) can substantially improve the quality and alignment of flow matching models for image and video generation. Methods such as Flow-GRPO and CPS cast the denoising process as a Markov Decision Process and apply PPO-style ratio clipping to enforce a trust region. However, we argue that *ratio clipping is structurally ill-suited for flow models*: the probability ratio between new and old policies is a noisy, single-sample estimate of the true policy divergence, leading to over-constraining in some regions of the trajectory and under-constraining in others. We propose **Flow-DPPO** (Flow Divergence Proximal Policy Optimization), which replaces ratio clipping with a divergence proximal constraint. A key observation is that the per-step policy in flow models is Gaussian, enabling *exact* and *cheap* computation of the KL divergence between old and new policies. Flow-DPPO employs an asymmetric divergence mask that blocks gradient updates only when they simultaneously move away from the trusted region and violate the divergence threshold. Experiments show that Flow-DPPO achieves higher rewards with better KL-proximal efficiency, alleviates catastrophic forgetting, promotes balanced multi-objective optimization, and enables stable multi-epoch training where ratio clipping degrades.

We develop a unified framework for flow-based reinforcement learning (RL) grounded in diffusion–flow duality and establish its theoretical foundations. The frameworks enables RL for deterministic flow matching inference by introducing diffusion stochastic inference dynamics that supports exploration while admitting deterministic deployment. We show that our frameworks subsumes several existing flow-based RL methods as special cases and inspire effective now methods. Then we provide theoretical foundations for the frameworks by establishing the correctness and performance transfer guarantee. Specifically, we prove that the policies optimized under stochastic dynamics close to deterministic dynamics at deployment when the pretrained model is well-trained. Moreover, we show that the policy improvement achieved under training-time SDE inference transfers to deployment-time ODE inference. Finally, we conduct experiments to validate our theoretical results.

We study token dynamics in deep Mamba models, with a particular focus on the attracting role of the first token. Unlike attention mechanisms, where token interactions are governed by positive scalar coefficients, Mamba dynamics involve full matrix-valued interaction coefficients with no positivity or definiteness guarantees. This structural difference makes Mamba dynamics substantially harder to analyze and prevents the direct application of existing techniques for attention-based models. To overcome this challenge, we introduce a new analytical framework based on a high-dimensional invariant box and a sharp exit-time argument, which together recover an effective form of sign-definiteness. Using this framework, we characterize the possible limit points of token trajectories and prove explicit exponential convergence rates. Our analysis reveals that the first token has privileged dynamics: its limiting direction attracts subsequent tokens, inducing token clustering and attention-sink phenomena in deep Mamba models. We further show that this attraction mechanism extends to broader sequence dynamics, including state-space models, causal attention, and classical time-series models. Experiments on Falcon-Mamba-7B validate our predictions, showing that concentration on the first token increases with depth and dominates in later layers.

Spiking Neural Networks (SNNs) are a promising framework for event-driven temporal processing. Prior work has improved temporal modeling through richer neuron dynamics and network-level mechanisms such as recurrence and delays, but it remains unclear how individual spiking neurons should specialize within a network. In this work, we introduce \textbf{FiTS}, a spiking neuron that factorizes temporal computation within each neuron into Frequency Selectivity (FS) and Temporal Shaping (TS). The FS module parameterizes each neuron's target frequency as the maximizer of its subthreshold magnitude response, while the TS module reshapes when frequency components contribute to membrane voltage accumulation through group-delay modulation. On auditory benchmarks where frequency selectivity and timing are central to the input structure, FiTS consistently improves over a plain Leaky Integrate-and-Fire (LIF) baseline in simple feedforward SNNs without recurrence or network-level delays, while remaining competitive with strong temporal SNN baselines. Beyond accuracy, the learned target frequencies and group-delay shifts provide interpretable neuron-level summaries of the frequency and timing organization learned within the network.

All-atom generative modeling of $3\mathrm{D}$ biomolecular complexes has emerged as the dominant paradigm for predicting the structure of proteins and protein-ligand systems. Generating structures at the atomic level of fidelity, however, typically requires expensive iterative diffusion rollouts, making both conventional deployment and inference-time search techniques computationally costly. In this paper, we introduce the $\textbf{De}$noiser $\textbf{C}$ofolding $\textbf{A}$ll-atom $\textbf{F}$lowmap (DeCAF) framework for distilling state-of-the-art all-atom cofolding models into all-atom flow maps that produce high-quality samples in only a few inference steps. We build DeCAF on a denoiser-based formulation of flow maps with endpoint losses that naturally support $\mathrm{SE(3)}$ rigid alignment, which we show is critical for training accurate models. We further derive a simple change of variables that lets DeCAF operate in the $\sigma$-space noise schedule of EDM-style architectures, enabling direct distillation from pretrained cofolding diffusion models. Equipped with DeCAF's flowmap lookahead, we introduce a purpose-built inference-time framework that improves sampling through reward-guided search. Empirically, DeCAF statistically improves over Boltz-1(x) in both accuracy (RMSD) and physical validity scores of protein-ligand poses at strict NFE budgets on the challenging Runs N’ Poses, while also showing a more optimal Pareto frontier across all inference compute budgets on PoseBusters.

As privacy, copyright, and safety requirements evolve, federated learning systems face a growing imperative to accommodate sequential unlearning requests. However, existing federated unlearning methods are designed for single-shot settings and their direct sequential application may lead to uncontrolled parameter-space overlap, causing cross-request interference that undermines prior forgetting. To address this, we propose the Spectral Orthogonality for Federated Continual unlearning (FedSOUL), a framework that reduces direct overlap among sequential unlearning updates. The core idea is to structure model updates in a shared orthonormal spectral basis, where each client is assigned a dedicated subspace in the frequency domain. Under this design, each unlearning request updates only its associated subspace, while all others remain unchanged, ensuring that sequential updates do not overlap in the allocated spectral coefficient space and thereby reducing cross-request interference. Beyond interference control, FedSOUL also improves efficiency by operating on sparse spectral coefficients, using FFT-based transforms for efficient computation and communicating only sparse coefficient updates. From a theoretical perspective, we relate continual unlearning to geometric conditions on update directions and show that our spectral construction satisfies update orthogonality by design, while functional preservation depends on bounded projected gradient leakage. Extensive experiments demonstrate that FedSOUL maintains stable unlearning–utility trade-offs across up to ten sequential requests and consistently outperforms competitive federated baselines. Our code is available at https://anonymous.4open.science/r/fedunlearning-1B92/.

World Action Models (WAMs) have emerged as a promising alternative to Vision-Language-Action (VLA) models for embodied control because they explicitly model how visual observations may evolve under action. Most existing WAMs follow an imagine-then-execute paradigm, incurring substantial test-time latency from iterative video denoising, yet it remains unclear whether explicit future imagination is actually necessary for strong action performance. In this paper, we ask whether WAMs need explicit future imagination at test time, or whether their benefit comes primarily from video modeling during training. We disentangle the role of video modeling during training from explicit future generation during inference by proposing **Fast-WAM**, a WAM architecture that retains video co-training during training but skips future prediction at test time. We further instantiate several Fast-WAM variants to enable a controlled comparison of these two factors. Across these variants, we find that Fast-WAM remains competitive with imagine-then-execute variants, while removing video co-training causes a much larger performance drop. Empirically, Fast-WAM achieves competitive results with state-of-the-art methods both on simulation benchmarks (LIBERO and RoboTwin) and real-world tasks, without embodied pretraining. It runs in real time with 110 ms latency, over 6$\times$ faster than existing imagine-then-execute WAMs. These results suggest that the main value of video prediction in WAMs may lie in improving world representations during training rather than generating future observations at test time.

Gradient descent has been of particular interest in modern machine learning beyond sole focus on optimization because some specific structures emerge through the optimization dynamics. Such behaviors result from optimization even though the learning objective does not explicitly encode the target structure, collectively called implicit bias, often preventing overparametrized models from fitting to spurious patterns. A typical instance is the max-margin implicit bias of a linear classifier, widely established for exponentially tailed loss functions. Even after having a given dataset separated, the parameter vector continues to evolve towards the max-margin direction asymptotically along the gradient descent dynamics. This phenomenon corroborates a frequent empirical observation of ``train longer, generalize better.'' However, the max-margin convergence is an asymptotic phenomenon, and what is worse, this asymptotic convergence rate is significantly slower than convergence in optimization. Even so, the parameter vector along gradient descent dynamics commonly correlates with the max-margin direction positively (though not exactly) within considerably fewer iterations than the asymptotic rate. By shedding another light on this classical yet profound problem, this work aims to understand the mechanism of this early-stage alignment phenomenon. Our theoretical results demonstrate that the parameter vector weakly aligns with the max-margin direction within $O(\exp(\exp(-\delta)))$ iterations, where $\delta>0$ is the permissible alignment error, which is shown to be tight. By tracking the radial and tangential flows, our proof operates on the alignment dynamics directly with dataset geometry and gets rid of the asymptotic expansion, which is a key insight to enabling faster weak alignment.

We introduce a novel Mutual Information (MI) estimator that fundamentally reframes the discriminative approach. Instead of training a classifier to discriminate between joint and marginal distributions, we learn a normalizing flow that transforms one into the other. This technique produces a computationally efficient and precise MI estimate that scales well to high dimensions and across a wide range of ground-truth MI values.

Air quality prediction at unmonitored locations is challenging because historical AQI observations are unavailable, while conventional spatiotemporal models largely rely on target histories to learn spatial dependencies. Meanwhile, widely available weather covariates provide useful station-specific evidence for target-absent inference. In this paper, we propose \textbf{ExtrapAir}, a lightweight weather-bridged framework for air quality inference at unmonitored locations. ExtrapAir preserves variable-specific weather patterns through per-variable weather encoding and learns adaptive weather and AQI correlations for spatial transfer. To support robust inference, we introduce \textbf{Activation Attention with Prior Inductive Biases}, which enables non-competitive spatial aggregation, incorporates geographic and semantic priors for station-pair guidance, and avoids full station-to-station attention computation for scalability. Extensive experiments on five real-world air-quality datasets against \textbf{32} baselines show that ExtrapAir reduces MAE by up to \textbf{15.24\%} overall and \textbf{16.01\%} on unmonitored stations, while cutting training time by \textbf{84.9\%} and memory usage by \textbf{76.23\%}, achieving a strong accuracy--efficiency trade-off.

Current evaluations of LLM safety predominantly rely on *severity-based taxonomies* to assess the harmfulness of models' responses to malicious queries. We argue that this formulation requires re-examination as it implicitly equates model output with realized harm while neglecting *Execution Likelihood*—the conditional probability of a threat being realized by a real-world actor given a model's response. In this work, we introduce **Expected Harm**, a metric that weights the severity of a jailbreak by its execution likelihood, modeled as a function of *execution cost*. Through empirical analysis of state-of-the-art models, we reveal a systematic *Inverse Risk Calibration*: models disproportionately exhibit stronger refusal behaviors for low-likelihood (high-cost) threats while remaining vulnerable to high-likelihood (low-cost) queries. Mechanistically, we first provide an analysis using linear probing, revealing that while refusal mechanisms linearly separate queries by severity, they do not use execution likelihood as an indicator for refusal. Then, we use Olmo-3's open-source staged checkpoints to trace the root cause of miscalibration to the Supervised Fine Tuning (SFT) training data distribution (82.8\% of safety examples at cost levels 0–1) and show that cost representations are weak from the SFT stage onward. Finally, we demonstrate that this miscalibration creates a vulnerability: by exploiting this property, we increase the attack success rate of existing jailbreaks by up to $2\times$.

In this paper, we introduce Exact Flow Linear Attention(EFLA), an exact-flow formulation of delta-rule linear attention. We show that the delta-rule update can be interpreted as an explicit Euler discretization of an underlying continuous-time system. EFLA replaces this first-order update with the exact closed-form flow. By exploiting the rank-1 structure of the dynamics matrix, both the matrix exponential and the input integral collapse to a simple update that preserves delta-rule linear attention's algebraic structure, parameter count, linear-time complexity, and chunkwise parallelism. This attention mechanism removes the Euler discretization error of the delta-rule dynamics without introducing additional parameters. Experiments on robustness tests, language modeling benchmarks, and the MAD synthetic benchmark show that EFLA improves stability under corrupted and high-energy inputs, reduces perplexity, and achieves stronger downstream performance compared to SSM and Euler-style baselines. These results establish exact-flow integration as a principled and scalable update mechanism for delta-rule linear attention.

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects (ω≤0.06), removing it causes small shifts (ω≤0.12) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points (ω≤0.31). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.

Entropy minimization (EM) is the dominant objective for test-time adaptation, yet its failure mode, model collapse, remains poorly understood. In this work, we show that distribution shifts can cause feature clusters corresponding to distinct classes in the model’s representation space to merge, while the decision boundary remains fixed. This induces a systematic skew in the predicted class distribution, referred to as prediction bias. Prediction bias refers to a shift in the predicted class distribution, with some classes overrepresented and others suppressed. We show that entropy minimization amplifies this prediction bias by tightening the existing clusters, reinforcing the incorrect groupings until all predictions collapse to a trivial solution. Next, to demonstrate the significance of prediction bias and mitigate it, we further propose Distribution Shift Bias Reduction (DSBR), a bias-correcting objective that specifically targets this failure mode by equalizing the contribution of each predicted class to the unsupervised entropy minimization loss. To study this failure mode, we design suitable adaptation settings using four medical‑imaging datasets and additionally evaluate on ImageNet‑C. We find that DSBR consistently stabilizes test‑time adaptation, prevents model collapse, and matches or outperforms the state‑of‑the‑art methods. Moreover, DSBR operates solely at test-time.