NeurIPS 2026 论文列表

Long-form reasoning and tool calling traces accumulate errors and stale content that anchor subsequent generations, a phenomenon known as \emph{context rot}. Existing scaffolds mitigate this with \textit{fixed-interval} compaction triggered at a token threshold, but such triggers are blind to trajectory structure and risk discarding partial results mid-derivation or mid-search. We propose SelfCompact, which pairs two inference-time elements: an inline \emph{compaction tool} the model invokes itself, and a lightweight \emph{rubric} specifying when to fire (a sub-task has resolved, or the trajectory is converging) and when to suppress (mid-derivation, or when stuck). Both are needed. The tool alone is unevenly used across open-weight models, often invoked at unhelpful moments or not at all; the rubric alone cannot act. Together, they elicit effective adaptive compaction without any fine-tuning. On competition math (IMO-Answerbench, HMMT Nov 25 / Feb 26) with four Qwen3 / Qwen3.5 models and agentic search (BrowseComp, BrowseComp-Plus) with three deployed agents, \method{} matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over a no-summarization baseline by up to 16.7 points on math and 5--9 points on agentic search at 30--70\% lower per-question cost. The result reframes \emph{when to compact} as a meta-cognitive capability that scaffolds, not weights, can supply.

We present Self-Cleaning Diffusion Models, a principled and effective framework for training diffusion models from heterogeneous data. In many domains, high-quality data is scarce or expensive to collect, while low-quality and out-of-distribution (OOD) data is abundant. Leveraging these ubiquitous samples is critical for scaling generative models. Still, principled techniques remain elusive, leaving practitioners to rely on simple heuristics such as high-quality finetuning or explicit quality conditioning. Self-Cleaning Diffusion Models bridges this gap by decoupling data correction from prior learning. We first train a transport map between the abundant OOD and the limited in-distribution samples. We then train an unconditional generative prior on a mixture of the few in-distribution samples and *noisy* versions of the transported points. Crucially, this noise injection prevents learning errors in the transport map from propagating to the generative prior. Theoretically, we demonstrate that transporting OOD samples towards the empirical proxy reduces the distance to the underlying distribution under mild assumptions. Empirically, we achieve state-of-the-art results across a diverse range of settings, from mitigating synthetic corruptions to increasing diversity and transcending dataset quality.

Developing generalist GUI agents capable of robust long-horizon execution remains a key challenge. While Reinforcement Learning (RL) offers a promising path for improvement, it is severely bottlenecked by the lack of reliable reward signals in open-ended environments. Ground-truth verification typically relies on expensive human annotation, which is prohibitively scarce at the scale required for training, while existing VLM-based judges frequently hallucinate success on superficial visual cues. This scarcity of trustworthy supervision fundamentally limits the scalability of current methods. In this work, we propose I-Judge, a novel framework for Self-Calibrated Reward Modeling. We leverage Inverse Dynamic Modeling (IDM) as a self-supervised training objective to learn the causal dynamics of GUI interactions from massive unlabeled trajectories, effectively bypassing the bottleneck of scarce human labels. We then introduce a runtime calibration mechanism that weights reward signals by the IDM's action consistency, filtering out spurious successes. Extensive end-to-end RL experiments on OSWorld, ScienceBoard and AndroidWorld demonstrate that our method significantly accelerates convergence and improves final agent performance. Notably, I-Judge demonstrates strong generalization to unseen domains. All the code, data and models will be made publicly available to foster further research.

The Forward-Forward (FF) algorithm trains networks layer-by-layer using a local "goodness function," yet the principles governing what makes an effective goodness function remain largely unexplored. We systematically explore the goodness-function design space and identify a unifying principle: the goodness function must be sensitive to the shape of neural activity, not its total energy. This principle is motivated by the observation that deep network activations follow heavy-tailed distributions and that discriminative information is often concentrated in peak activities. We propose two complementary families: selective functions (top-k, entmax-weighted energy) that measure only peak activity, and shape-sensitive functions (excess kurtosis / "burstiness" and higher-order moments) that reward heavy-tailed distributions via scale-invariant statistics. Combined with separate label-feature forwarding (FFCL), controlled experiments across 13 goodness functions, 5 activations, 6 datasets, and three continuous sweeps each tracing a characteristic inverted-U yield 89.0% on Fashion-MNIST and 98.2±0.1% on MNIST (4x2000)—a +32.6pp gain over SoS—with consistent improvements across all benchmarks (+72pp USPS, +52pp SVHN). The scale-invariant nature of burstiness makes it particularly robust to magnitude shifts across layers and datasets. Code is available at [https://anonymous.4open.science/r/ff-selectivity-shape](https://anonymous.4open.science/r/ff-selectivity-shape).

Selective state-space models such as Mamba replace fixed linear state transitions by input-dependent transitions, yielding recurrent sequence models that retain linear-time inference while adapting their memory dynamics to the current token. We show that even a simple form of this selective memory update has universal approximation power. Specifically, for every compact input domain, finite horizon, and continuous causal sequence-to-sequence map, a clocked diagonal selective state-space scan with strictly contractive gates and a linear-in-state readout uniformly approximates the target map. The proof is constructive: input-dependent diagonal propagation generates tensor-product partition features over prefixes, and the linear-in-state readout combines these features with sampled target values. Thus the high-order interactions required for universality are produced inside the selective scan itself, rather than by attention, nonlinear hidden-state dynamics, or a universal multilayer perceptron applied after a sequence encoder. These results identify selectivity as a mathematically sufficient mechanism for expressive linear-time state-space sequence modeling. Selectivity lets the model decide, token by token, which parts of the past should be preserved, faded, or combined with the present. The universal approximation theorem shows that this adaptive use of memory is sufficient to reproduce any well-behaved causal sequence rule over a fixed horizon.

While large language models (LLMs) are trained to align with human values, their generations may still violate safety constraints. A growing line of work addresses this problem by modifying the model’s sampling policy at decoding time using a safety reward. However, existing decoding-time steering methods often intervene unnecessarily, modifying generations that would have been safe under the base model. Such unnecessary interventions are undesirable, as they can distort key properties of the base model such as helpfulness, fluency, style, and coherence. We propose a new test-time steering method designed to reduce such unnecessary interventions while improving the safety of unsafe responses. Our approach filters tokens using a value-based safety criterion and provides an explicit bound on the probability of false interventions. A single threshold hyperparameter controls this bound, allowing practitioners to trade off higher rates of unnecessary intervention for better output safety. Across multiple datasets and experiments, we show that our value-filtered decoding method outperforms existing baselines, achieving better trade-offs between safety, helpfulness, and similarity to the base model.

Group-relative RL training (GRPO) samples a small group of parallel rollouts for every training prompt and uses their within-group reward spread to compute per-trajectory advantages. In agentic environments each rollout is a long multi-turn dialogue with one LLM call per step, so this multi-sample multiplier dominates the total training cost. When every rollout of a prompt ends with the same reward, the group has zero reward variance and contributes no gradient, so the extra rollouts add no information; such groups are common in practice (typically around 40% of all groups), so the wasted-compute fraction is substantial rather than marginal. Existing methods filter such groups at the prompt level, either after their rollouts are paid for or before any rollout begins, but both decide without using information that becomes available during the rollout itself. We instead ask whether the in-group divergence between the partial trajectories at an intermediate step can already predict that the group will be zero-variance: when the parallel rollouts have already converged on the same action prefix, the group is on track to produce a single reward, and we can stop early. We propose a one-parameter gate that stops a group when the mean pairwise prefix edit distance between its partial action sequences falls below a threshold. On a 60-iteration on-policy GRPO run on ALFWorld with Qwen2.5-7B, averaged over four random seeds, the gated arm finishes 10.7% faster in wall-clock (bootstrap 95% CI excludes 0) and shifts held-out success rate on 50 unseen tasks by +2.5 pp, with the held-out gain tracing to a measurable reduction in zero-advantage gradient-batch dilution.

Reinforcement learning from verifiable rewards improves reasoning by reinforcing sampled solutions that receive positive outcomes, but group-relative methods such as GRPO become silent on hard prompts where every sampled rollout is wrong. These zero-reward prompts are often the most informative failures: each comes with a verified reference solution, yet uniformly imitating the full trace mixes core reasoning decisions with routine algebra, formatting, and surface wording. We propose **Selective Off-Policy Reference Tuning with Plan Guidance (SORT)**, an auxiliary repair update that leaves GRPO rollout generation unchanged. For each failed prompt, SORT extracts a reference-derived reasoning plan and scores every reference token twice under the same model, once with only the problem and once with the plan added to the context. Tokens whose probabilities rise under plan conditioning are treated as structurally informative and receive larger Dynamic Fine-Tuning updates, while tokens outside the model's support remain controlled by the base probability. We formalize this mechanism with a ground-truth plan model and prove that the SORT weight approaches an oracle structural-token weight as the extracted plan better approximates the true plan-conditioned distribution. Across three instruction-tuned backbones and eight in-distribution and out-of-distribution reasoning benchmarks, SORT consistently improves over GRPO and guidance-based baselines, with the largest gains on the weakest model where zero-reward failures are most frequent.

Rotation invariance is a fundamental requirement across many computer vision tasks. Historically, this inductive bias has been encoded through hand-crafted rotation-invariant representations. These are compact, interpretable, and fast to compute, but they come at the cost of descriptive power. More recently, architectures achieve inductive bias through learned representations. These are highly descriptive and achieve strong empirical performance, at the cost of efficiency and interpretability. In this work, we propose an alternative at the intersection of both paradigms. We introduce the selective disk bispectrum (SDB), a complex-valued rotation-invariant vector that preserves all information about the image except its orientation. Our key theoretical contributions are the selective disk bispectrum, its inversion, its (reduced) spatial and computational complexities (compared to the full disk bispectrum), and its expectation and variance under noise. Furthermore, we propose a numerical SDB approximation and provide theoretical guarantees for its accuracy and rotation invariance. Empirically, we validate SDB's invariance and robustness to noise classification tasks. We test our reconstruction algorithm on multi-reference alignment of rotated images.

Medical visual question answering is a safety-critical decision problem in which abstaining on inputs the image cannot support matters as much as answer accuracy. Existing inference-time agentic methods address hallucination by adding a verification step, but they collapse the final decision at a single language-model judge that integrates all evidence in one prompt and is systematically biased toward declaring inputs answerable. We propose PICV (Parallel Independent Claim Verification), which reformulates selective medical VQA as independent claim verification with transparent aggregation: the question is decomposed into a small set of typed visual claims, each is verified by an isolated prompt with no access to the other verdicts, and the verdicts are combined by a deterministic rule rather than another model, yielding an answer/abstain decision together with a typed unanswerability reason. We evaluate PICV efficiently on an automated answerability benchmark we construct over two public radiology datasets via three reproducible corruption procedures, requiring neither heavy VLM judging nor human annotation at evaluation time. Across four VLM backbones, PICV consistently outperforms strong prompting and agentic baselines; ablations attribute the gains specifically to per-claim isolation and rule-based aggregation rather than to agentic decomposition alone.

Radar-camera depth estimation must turn an ultra-sparse, all-weather, metric radar signal into a dense per-pixel depth map. Existing methods --- concatenation, confidence-aware gating, sparse supervision, graph-based extraction --- combine radar and image features outside the backbone's sequence operator, and even cross-modal Mamba variants leave the selection mechanism itself unimodal. We argue that the selection mechanism is the right place for radar to enter. We introduce Radar-Modulated Selection (RMS), a minimal and principled way to inject radar into Mamba's selective scan: radar modulates the scan from within, adding zero-initialised perturbations to the step size $\boldsymbol{\Delta}$ and readout $\mathbf{C}$ while leaving the input projection $\mathbf{B}$ and state dynamics $\mathbf{A}$ image-only. The construction is exactly equivalent to a pretrained image-only Mamba at initialisation, ensuring radar only influences the model where it improves accuracy. Two further properties follow that out-of-scan fusion cannot offer: linear-cost cross-modal coupling at every recurrence step, and a natural fallback to the image-only backbone when radar is absent. We deploy RMS in a Multi-View Scan Pyramid (MVSP) that matches the fusion operator to radar's spatial reach at each scale. SemoDepth achieves state-of-the-art performance on nuScenes, reducing MAE by 34.0%, 29.9%, and 29.9% over the previous best at 0-50, 0-70, and 0-80 m, while attaining the lowest single-frame latency (26.8 ms). A further ablation shows that out-of-scan feature blending adds no accuracy on top of RMS, providing empirical validation that in-scan selection can replace out-of-scan fusion.

Today, verifier-guided best-of-$N$ inference is a common way to spend test-time compute in language modeling, code generation, and alignment: generate $N$ candidates, score them with a verifier or reward model, and deploy the highest-scoring output. This paper argues that the relevant reliability object is not average verifier accuracy but \emph{selected-tail reliability}, namely the behavior of the output chosen after optimizing over $N$ noisy public scores; we derive exact selected-rank exposure laws for best-of-$N$ selection and finite-verifier resource laws showing that selected public optimism scales at the $\sqrt{\log N/m}$ regime under $m$ units of verifier evidence. We then separate public optimism from hidden utility harm, showing that public-score inflation alone is not a harm claim and that harm or uncertifiability requires public-hidden mismatch, tail dominance, count imbalance, or an unresolved selected-tail regime. Building on these results, we introduce Tail-Certified Compute Caps, a conservative procedure that certifies candidate budgets or refuses certification using selected-tail intervals, verifier-resource penalties, mismatch envelopes, and feasibility gates. In exact-verifier/public-hidden experiments, the theory predicts selected false-positive exposure and supports resource-aware cap decisions over an all-attempted denominator, while non-code candidate tables with pre-existing public human annotations serve as secondary scoped $K\ge16$ consistency checks under proxy/model-prior/verifier-style scores across Arena55K and Stanford SHP. Overall, the paper provides a reliability framework for verifier-guided inference-time scaling: candidate budgets should be justified by evidence about the selected tail that is actually deployed, not only by average verifier validation.

We study optimistic bilevel optimization when the lower-level problem has a non-isolated manifold of minimizers. In this setting, the hyper-objective may be non-differentiable because the upper-level criterion must choose among multiple lower-level solutions. Under a local Polyak--{\L}ojasiewicz (P{\L}) condition, we show that differentiability does not require the lower-level solution set to be a singleton: uniqueness of the optimistic selection is sufficient. This yields an explicit pseudoinverse-based hyper-gradient formula extending the classical singleton-minimizer result. We further characterize the regularity of the hyper-objective: non-degeneracy of the selected minimizer along the solution manifold yields local smoothness, while failure of uniqueness can create many non-differentiable points and failure of non-degeneracy can destroy all positive H\"older regularity of the hyper-gradient. Motivated by this theory, we propose HG-MS, a select-then-differentiate method combining explicit optimistic selection with efficient pseudoinverse-based hyper-gradient computation. Despite the nonconvex nature of optimistic selection over the lower-level solution manifold, we show that HG-MS converges to a stationary point of the optimistic objective with complexity governed by the intrinsic dimension of the solution manifold rather than its ambient dimension. Empirically, we test a practical variant of HG-MS for matched-budget LLM source reweighting. This variant preserves the select-then-differentiate principle and obtains the best GSM8K/MATH scores across the tested backbones, along with competitive or best MT-Bench instruction-following results.

Prompt optimizers invest heavily in how to generate better prompts, yet pay almost no attention to which examples they use to judge them. This evaluation subset directly shapes every feedback signal that the optimizer receives, while existing methods either fix it before optimization begins (principled but agnostic to the evolving prompt population) or adapt it heuristically (flexible but unstable). We bridge this gap by constructing an online adaptive testing problem: Prompts are examinees, training examples are test items, and the scheduler selects items that best discriminate among the strongest candidates. We introduce POES (Prompt-Aware Online Evaluation Scheduling), whose monotone submodular objective combines discrimination, coverage, and bounded subset updates, yielding a $(1-1/e)$ cold-start guarantee and a warm-start tracking bound under bounded inter-round drift. Across a 35-task APO suite plus 57 MMLU subjects and 6 optimizer-model configurations, POES achieves the highest mean accuracy in every configuration, improving over the best baseline by +3.0 pp to +7.8 pp. Notably, the achieved gains concentrate on tasks where baselines still have headroom to improve, vanishing where methods saturate, suggesting that what you evaluate on matters most precisely when there is a signal to extract.

We present Seirênes, a self-play RL framework that transforms contextual interference from a failure mode of LLM reasoning into an internal training signal for co-evolving more resilient reasoners. While RL with verifiable rewards has significantly advanced reasoning capabilities, models can still exhibit fragility when encountering non-idealized contexts: scenarios characterized by superfluous information, tangential instructions, or incidental correlations that differ from the clean distributions typical of standard benchmarks. Seirênes harnesses this vulnerability through a parameter-shared and adversarial self-play loop. Within this framework, a single model is trained to both construct plausible yet distracting contexts that expose its own reasoning blind spots, and solve problems by discerning the essential task from these perturbations to recover the core underlying logic. By pitting these competing objectives against each other, Seirênes compels the model to move beyond superficial pattern matching and anchors its capabilities in robust underlying reasoning. This continuous interaction sustains an informative co-evolutionary curriculum as the model improves. Across seven mathematical reasoning benchmarks and model scales from 4B to 30B, Seirênes achieves average gains of +10.2, +9.1, and +7.2 points. Besides, distracting contexts produced by the 4B Seirênes model reduce the accuracy of top-tier closed-source models (GPT and Gemini) by roughly 4--5 points, revealing Seirênes' general ability to uncover reasoning models' blind spots. Our code will be available.

3D vision-language segmentation aims to segment target objects in 3D scenarios according to the linguistic instructions and visual observations. Prior art heavily relies on the coarse superpoint representation to reduce the computation complexity , which suffers from poor segmentation quality and messy object boundaries. In this paper, we propose the SEGment-And-select (SEGA3D) paradigm for 3D vision-language segmentation that directly operates on the fine-grained visual information and is free from the superpoint dependency. Specifically, we first leverage a mask candidate generator to provide fine-grained categorical mask candidates, substantially improving the quality of candidate masks over the superpoint counterparts. Then, a Large Language Model (LLM) is utilized to generate the semantic and spatial information based on the linguistic description and visual features. The LLM output and visual features are fed to the Semantic-Spatial Selector (SSS) to produce the top-ranking mask candidates. Eventually, the Loopback Verification Module (LVM) is designed to yield the segmentation mask from the selected candidate masks. Our SEGA3D attains competitive performance on ScanRefer, ScanNet and Matterport3D benchmarks. Notably, our SEGA3D surpasses the top-performing counterpart by 8.3 mIoU and 5.3 mIoU on ScanNet and Matterport3D, respectively. Codes will be available upon publication.

Part-level 3D assets are essential for editing, reassembly, and interaction, yet recovering such structure from a single image remains challenging due to occlusion, ambiguous boundaries, and the need for coherent multi-part reasoning. Existing approaches struggle to achieve both controllable part-level generation and coherent multi-part structure, as part identity and spatial allocation are typically inferred implicitly. We present Seg3DParts, a segmentation-grounded framework for controllable part-level 3D generation from a single image. By treating segmentation as an explicit grounding signal, our method defines part identity during generation, enabling each component to be anchored to a corresponding image region. To ensure coherent assemblies, we introduce structured cross-part interaction that allows components to exchange global context throughout the generative process. As a result, Seg3DParts directly generates well-aligned part meshes in a shared canonical space without post-hoc alignment, supporting flexible and controllable decomposition. We further introduce PartObjectNet, a large-scale dataset with over 200K objects and 1M annotated parts. Experiments demonstrate that Seg3DParts achieves superior geometry quality, cross-part coherence, and part-level controllability over existing methods.

Generalized category discovery (GCD) becomes substantially more challenging when the number of novel classes is unknown. We argue that, after the backbone is fixed, this stage should be treated as seen-constrained model-order selection rather than as an auxiliary clustering detail. The selector must estimate the class count without novel labels while remaining consistent with the labeled seen categories available in the standard GCD setting. We instantiate this view by constructing a nested merge path from a seen-constrained overclustering and selecting the coarsest near-optimal cut under observable surrogate decision risks. These risks combine seen-class consistency, cross-seed stability, unlabeled-side compactness, and a tiny-cluster degeneracy penalty. The resulting protocol separates class-count error, downstream clustering quality, and robustness. Where an aligned public unknown-$K$ baseline is available, namely on SelEx, the selector reduces class-count error on all four fine-grained datasets, including a reduction from 342 to 15 classes on Herbarium-19. Broader-domain diagnostics further distinguish ImageNet100, where the default selector can underestimate the class count toward the seen-class regime, from CIFAR100, where the selector landscape exposes an accuracy-count operating-point trade-off. The mixed accuracy outcomes reveal distinct count-, path-, and representation-level bottlenecks, supporting unknown-$K$ selection as a decision layer that should be evaluated separately from representation learning.

What does it mean to create a new concept, rather than retrieve a familiar one? Diffusion models sampled many times at the same prompt vary only within the narrow stylistic envelope set by a fixed prompt embedding. We propose an operational definition of creativity as the generation of concepts that are initially unfamiliar yet rapidly learnable to an adaptive observer, and formalize it as a bilevel optimization between a Creator that generates and an Appraiser that adapts: the Appraiser's improvement under a brief inner-loop adaptation provides the reward signal that the Creator maximizes. We call this framework CAMEL (Creator-Appraiser Meta-Learning), and instantiate it on MNIST with an autoencoder Appraiser and on natural images with a CLIP Appraiser based on a low-rank adapter over the text projection. CAMEL produces outputs that lie outside the basin reachable by vanilla diffusion sampling at any sample count, while remaining recognizable as the prompt's class to an off-the-shelf observer.