NeurIPS 2026 论文列表

Lightweight vision-language models perform competitively on standard benchmarks yet fail systematically in dense-scene reasoning, where multiple objects, attributes, and relations must be jointly grounded and resolved through multi-step inference. Such capability is critical for real-world applications where models must reliably interpret cluttered environments. Yet existing training signals provide no explicit grounding between reasoning steps and the underlying visual entities and relations, leaving lightweight models free to generate fluent but visually unanchored reasoning chains. To address this gap, we first introduce DRBench, a benchmark of 14,573 questions across 2,943 images, organized into five task categories spanning three progressive reasoning layers. Building on DRBench, we propose DRScaffold, a supervised fine-tuning framework that decomposes the supervision target into four causally ordered stages, enforcing grounded reasoning without architectural modification. Experiments on three lightweight VLMs demonstrate substantial gains on DRBench while preserving or improving performance on general-purpose benchmarks. Notably, Qwen2.5-VL-3B trained with DRScaffold surpasses the frozen Qwen2.5-VL-32B on DRBench, demonstrating that structured supervision can substitute for a significant portion of model scale in dense-scene reasoning.

Single-stage fine-tuning is pervasive, yet catastrophic forgetting persists when the source pre-training data is inaccessible. Data weighting affords a natural lever to modulate sample influence. Yet existing weighting methods lack a principled mechanism to explicitly arbitrate knowledge retention against new-task adaptation, and provide no unified differentiable objective to accommodate auxiliary signals such as sample diversity and sample characteristics. This is because current schemes privilege one goal over the other, and no explicit gradient-compatible encoding exists for such signals. To address this, we propose $\textbf{DRIVE}$, a principled and efficient weight assignment framework. It derives per-sample weights via a unified differentiable objective where three coupled signals co-evolve: Shapley additivity splits contribution into anti-forgetting and adaptation components, which are reshaped by a covariance-based diversity term and stabilized by a sample-characteristic prior. Far from a mere combination, this yields an end-to-end gradient-based scheme balancing stability against plasticity. Theoretically, we prove that DRIVE enjoys superior stability-plasticity guarantees over existing weighting schemes and enhances parameter estimation robustness. Empirically, it attains superior overall performance, faster convergence, and stronger robustness. Code at https://anonymous.4open.science/r/DRIVE-158.

World models are increasingly used to reduce environment interaction when post-training vision-language-action (VLA) policies with reinforcement learning. Yet they are typically trained to predict pixels one step ahead, even though their role here is not merely to render visually faithful videos but to produce learning signals from imagined rollouts that improve the policy. We show that this mismatch is substantial: lower pixel MSE can produce worse policy gradients, and a world model trained for policy utility can yield better downstream policies despite 17\% higher pixel MSE on the next frame. We propose the Downstream-Return Imagination Learning Loop (DRILL), a bilevel framework that updates the world model to maximize the policy's return after an inner GRPO update on imagined rollouts. DRILL has two instantiations. DRILL-VWI is a closed-form surrogate, requiring no additional simulator rollouts, that reweights prediction errors by policy visitation, the group-relative GRPO advantage, and model confidence. DRILL-IMG is a full meta-gradient method that differentiates through the inner update using truncated Hessian-vector products. We show that DRILL-VWI recovers the local one-step component of the DRILL-IMG meta-gradient under score-compatibility and small-residual assumptions. Across five manipulation simulators and two VLA backbones, DRILL-IMG raises mean success rate by 13.1 points over WMPO and 7.9 points over RLVR-World, and complements RLVR pretraining rather than competing with it. More broadly, we believe video world models for VLA reinforcement learning can be productively trained by the policy gradients they produce, not only by pixel-level prediction.

Diagram question answering (DQA) requires models to interpret structured visual representations such as charts, maps, infographics, circuit schematics, and scientific diagrams. Recent vision-language models (VLMs) often achieve high answer accuracy on these tasks, yet correct answers do not guarantee that models ground their reasoning in the diagram regions that support the prediction. Models may instead rely on textual correlations or dataset artifacts without identifying the visual evidence required to verify the answer. This limitation prevents reliable evaluation of diagram reasoning and reduces interpretability. We introduce **DRAGON**, a benchmark for evaluating evidence-grounded visual reasoning in diagrams. Given a diagram, a question, and the correct answer, a model must predict bounding boxes that correspond to the visual elements required to justify the answer. These evidence regions may include answer-bearing components, textual labels, legends, axes, connectors, and other supporting structures involved in the reasoning process. The DRAGON dataset contains 11,664 annotated question instances from six diagram QA datasets: ChartQA, Circuit-VQA, InfographicsVQA, MapIQ, MapWise, and AI2D, with a 2,445-instance test set carrying human-verified evidence annotations and a standardized evaluation framework. Our evaluation on recent VLMs reveals that even Claude Opus 4.6 achieves a Grounding IoU of only 23.8% on AI2D, showing that faithful visual grounding remains an open challenge. DRAGON supports future research on models that ground their predictions in visual evidence.

Differential privacy (DP) has a wide range of applications for protecting data privacy, but designing and verifying DP algorithms requires expert-level reasoning, creating a high barrier for non-expert practitioners. Prior works either rely on specialized verification languages that demand substantial domain expertise or remain semi-automated and require human-in-the-loop guidance. In this work, we investigate whether large language models (LLMs) can automate DP reasoning. We introduce DPrivBench, a benchmark in which each instance asks whether a function or algorithm satisfies a stated DP guarantee under specified assumptions. The benchmark is carefully designed to cover a broad range of DP topics, span diverse difficulty levels, and resist shortcut reasoning through trivial pattern matching. Experiments show that while the strongest models handle textbook mechanisms well, all models struggle with advanced algorithms, revealing substantial gaps in current DP reasoning capabilities. Through further analytic study and failure-mode analysis, we identify several promising directions for improving automated DP reasoning. Our benchmark provides a solid foundation for developing and evaluating such methods, and complements existing benchmarks for mathematical reasoning.

The remarkable scalability of Transformers has expanded their application to 3D computer vision, where camera-aware positional encoding is crucial for providing spatial cues in multi-view geometry. Recent advancements have established the practice of using camera parameters---such as extrinsics or projection matrices---as relative positional encoding into the query, key, and value vectors of the attention mechanism. However, when scaling up the training recipe of novel view synthesis (NVS) models with the camera-based positional encoding, we observe a significant issue: model performance stagnates in the late stages of training. In this paper, we investigate the cause of the performance bottleneck when scaling up and demonstrate that storing rotation and translation given by the positional encoding in the same dimensions of the value vector causes indeterminacy in their independent identification, hindering training scalability. To address this, we propose Decoupled Pose Positional Encoding (DPPE), a novel camera-based positional encoding that explicitly decouples rotation and translation. Extensive evaluations on NVS tasks demonstrate that DPPE enables stable long-term training even in scaled-up training setup. Furthermore, it exhibits superior generalization performance in extrapolation settings, such as handling an increased number of viewpoints and zoom-in scenarios.

Decentralized training of deep learning models is widely used to enable data privacy and on-device learning over networks. In realistic scenarios, heterogeneity across clients' local data poses an optimization challenge and can severely degrade test accuracy, especially on sparse communication graphs. Building on recent advances in centralized primal averaging, we propose Decentralized Primal Averaging (DPA), a decentralized learning algorithm that combines primal averaging with a local quasi-global momentum estimate. DPA maintains three coupled sequences for optimization, gradient evaluation, and output averaging, while using consecutive mixed iterates to construct a network-aware momentum direction without transmitting an auxiliary tracking variable. We prove a nonconvex stationarity guarantee for DPA's averaged output model and establish a linear-speedup result up to network- and heterogeneity-dependent residual terms. We also introduce DPA-1G, a one-gossip-per-iteration variant that matches the communication budget of DSGD-style baselines. Experiments following the GUT decentralized image-classification benchmark show that DPA outperforms DSGD, gradient tracking, QG-DSGDm, momentum tracking, and QG-GUTm across datasets, architectures, topologies, and heterogeneity levels, with the largest gains under high heterogeneity. DPA-1G matches or exceeds the one-gossip baselines at the same per-iteration communication budget.

Evolution strategies optimize through fitness evaluations, but standard ES exposes data-dependent fitness values. We introduce DP-EGGROLL, a central-DP mechanism for EGGROLL-style population optimization with decomposable supervised losses. Each private step forms per-example candidate-loss vectors, mean-centers each row across candidates, clips the centered row, averages with a fixed nominal denominator, and adds calibrated Gaussian noise; reward shaping and the ES update are post-processing. Clipping gives the $C/m$ add/remove sensitivity bound, while centering removes common-mode loss before clipping. We evaluate 10 public tabular benchmarks across five privacy budgets, using 10 final seeds and equal 32-configuration HPO budgets for fast DP-AdamW, scalar DP-ZO, centered DP-EGGROLL, and independently tuned uncentered DP-EGGROLL. Each selected final run is accounted as an individual $(\varepsilon,10^{-5})$-DP run; multi-configuration HPO is a benchmark protocol, not a single-run deployment guarantee. On tabular classification, centered DP-EGGROLL is practically non-inferior to fast DP-AdamW on 22/25 AUROC endpoints and is faster per step in all 25 classification settings. Centering improves over tuned uncentered DP-EGGROLL on all 25 classification AUROC endpoints. Neural experiments include end-to-end small MLPs and a low-rank adapter stress test; both reinforce that centered population-vector privatization carries more signal than scalar DP-ZO. Regression is more heterogeneous, with centered DP-EGGROLL practically non-inferior on 14/25 RMSE endpoints.

Asynchronous reinforcement learning (RL) has become a critical paradigm for accelerating large-scale LLM post-training in industrial settings, yet it faces a structural \emph{long-tail dilemma}: rollout efficiency is bottlenecked by a small number of the longest trajectories, which are precisely the most valuable ones for RL training. Existing approaches alleviate this dilemma at the cost of either system overhead (e.g., re-prefill in partial-rollout methods) or algorithmic compromises (e.g., discarded long trajectories in replication-based methods). We trace these tradeoffs to a common assumption: the rollout cluster serves a single policy version at any moment. We propose \textbf{DORA} (\textbf{D}ynamic \textbf{OR}chestration for \textbf{A}synchronous Rollout), which breaks this assumption by maintaining multiple policy versions concurrently within the rollout cluster. DORA combines three mechanisms: \emph{multi-version streaming training} that decouples trajectory completion from batch boundaries, a centralized \emph{load-balancing orchestrator} that re-partitions resources across versions, and \emph{zero-re-prefill migration} that transfers KV-Cache directly across same-version instances. Experiments on open-source benchmarks show that DORA achieves up to $2.12\times$ end-to-end throughput improvement and $8.2\times$ rollout-stage acceleration over synchronous training while preserving convergence parity. \textbf{In real-world production deployment} with thousands of accelerators, DORA achieves up to $6.2\times$ rollout speedup and produces competitive open-source LLMs.

We present DMax, a new paradigm for efficient Diffusion Language Models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On one single H200 GPU using the SGLang framework, our 16B model achieves an average of 955 TPS at a batch size of 1.

Multimodal Large Language Models (MLLMs) have shown strong multimodal instruction-following ability, but adapting them to diverse visual-language domains typically assumes centralized data access and costly joint training. This is restrictive when data is distributed across private, domain-specific, or permission-limited clients. To this end, we propose DISTMOE, a mixture-of-experts (MoE) approach for distributed visual instruction tuning. In each layer of the language decoder it augments the public feedforward network (FFN) with a client-specific private FFN expert, with the goal to acquire domain-specific knowledge. However, independent expert training causes the private FFNs to learn representation of different scale and magnitudes, making merging the experts difficult. To reduce client-specific drift, we introduce a public-anchored expert composition stage that updates only routers and lightweight private projection adapters on a mix of local client data and public data, via an isotropic regularization loss, therefore making it rehearsal-free. During inference, DISTMOE performs modular routing over public and private experts, enabling token-wise domain composition without explicit domain labels. Experiments across diverse visual-language benchmarks show that DISTMOE enables flexible expert reuse, effective domain adaptation, and competitive performance while preserving modular control over client-specific knowledge.

Recent advances in reinforcement learning have made verifiable rewards a central post-training paradigm for improving multimodal reasoning. However, methods such as GRPO remain data- and compute-inefficient, often spending rollout and annotation budget on prompts that provide little learning signal. We propose DISCOVER, an online data-discovery framework for budgeted multimodal GRPO that enables vision-language models to dynamically identify prompts that are most useful for the current policy. DISCOVER first organizes the multimodal sample pool into semantic clusters and queries a small set of density-diverse representative samples to obtain initial reward observations. During training, it estimates the utility of unannotated samples by transferring sparse reward feedback from queried samples. Since sample utility changes as the policy evolves, DISCOVER dynamically updates these estimates with recency-weighted rollout observations. It then selects only high-utility prompts for annotation using a policy-conditioned score that accounts for reward variance and difficulty, and continues utility-based replay once the discovery budget is exhausted. Experiments on ViRL39K and VLAA-Thinking with Qwen3-VL-2B show that DISCOVER substantially improves multimodal mathematical reasoning under a strict 10\% unique-sample budget, outperforming budget-matched random selection, hard-example sampling, full-pool GRPO under the same training-step budget, and online selection baselines. Ablations further validate the importance of semantic coverage, online reward transfer, and stable discovery cadence, establishing DISCOVER as a practical framework for data-efficient multimodal RL post-training.

LLM-based coding agents have shown strong capabilities on real-world software testing tasks, yet existing evaluations concentrate on unit test generation, leaving system testing unexplored. System testing is both practically important for software quality assurance and inherently difficult, as it demands repository-level holistic understanding and long-horizon reasoning over program behavior. In this paper, we propose to measure LLM agents' system testing capabilities through its core problem, Directed Input Synthesis (DIS): generating system-level inputs that exercise specified code branches. We introduce DIS-Bench, a benchmark built via an automated pipeline based on random testing that selects target branches that are both reachable and challenging. Because DIS-Bench is constructed from execution behavior rather than human-authored issues or patches, it is scalable and less susceptible to contamination from existing training corpora. DIS-Bench contains 8267 tasks drawn from 37 real-world repositories. We evaluate 9 mainstream LLMs on DIS-Bench-Lite under both BM25 retrieval and coding agents, including OpenHands, Mini-SWE-Agent, and Claude-Code. The results show that DIS is highly challenging: the best open-weight configuration resolves only 28.1\% of targets at pass@1, the best frontier configuration (57.8\%) still leaves over 40\% unsolved, and 23.1\% of targets are unsolved by every configuration in every attempt. We further perform a manual inspection to characterize the dominant failure modes, revealing reasoning bottlenecks of LLMs on system testing tasks.

Open-vocabulary 3D Gaussian segmentation typically binds semantic features to per-Gaussian state via per-scene gradient training (1–4 h) or the more recent closed-form distillation. Both commit each primitive to a single feature; therefore, boundary Gaussians spanning multiple 2D instances are forced into a hard decision that degrades retrieval accuracy and yields unstable mask boundaries at render time. We present DIGS, a training-free framework with two stages, class-agnostic map- ping and multi-modal encoding, that decouples instance discovery from linguistic naming. At the Gaussian level, our distribution-informed representation maintains an explicit top-M label distribution per primitive, updated by a tempered Dirichlet rule; preservation of multi-peaked posteriors at boundary primitives allows a late-argmax renderer to recover clean instance boundaries. A bidirectional state machine aggregates these distributions into a persistent class-agnostic instance ontology. The instance-level encoding stage then fuses the CLIP visual embedding of each instance’s isolated 3D footprint with the CLIP-text embedding of an LLM-generated attribute description, disambiguating visually similar instances. DIGS maps a scene in 1–3 minutes on a single A100 and serves 3D-native retrieval in 5.4 ms, reaching 71.95% mIoU on LERF-OVS, surpassing the strongest training-free baseline SFS and the strongest training-based baseline LaGa, and setting new state-of-the-art performance on LERF-Mask (91.71%) and 3D-OVS(96.45%). Codes are at: https://anonymous.4open.science/r/DIGS-C4EF.

Fluent guidance is not the same as useful intervention. LLM tutors are typically trained to generate the next teacher utterance, implicitly assuming that every student turn warrants a response. However, our experiments indicate that this conflates tutoring capability (what to say) with intervention necessity (whether to say it). We introduce DICE, a framework that decouples intervention decisions from response generation by first selecting an explicit pedagogical action. To calibrate this action selection policy, we define Intervention Value (IV), a rollout-grounded counterfactual metric that compares each action against non-intervention. IV shows that many prescribed interventions provide little or no marginal benefit. We further introduce DICE-Bench, a multi-variant math tutoring benchmark with skill-preserving problem variants for session-level evaluation. Using IV-weighted and KL-regularized policy optimization, DICE learns to intervene selectively while preserving tutoring effectiveness. In simulated tutoring sessions, DICE reduces the over-intervention rate to near zero while guiding students to correct solutions in approximately 3-4 fewer turns on average than existing Socratic tutoring baselines.

We introduce DIAL (Directional Intervention, Asymptotically Limited), a parameter-efficient adapter that adds a bounded, monotone, scalar-controlled residual $\delta(s,c) = \sigma(g) \odot \tanh(c/T) \odot \tanh(Dz)$ with $z = D^\top \phi(s)$ to a base language model's hidden state. The architectural form yields a closed-form upper bound $\lVert \delta \rVert_\infty \le \sigma(g_{\max})$ on the residual $L_\infty$ norm, computable from model weights alone. Across 5 base models (3.8B to 70B parameters, 4 architecture families; one checkpoint per family, with 5-seed csweep variance at Llama-3.1-8B), the residual saturates at the predicted bound with mean ratio $1.000 \pm 0.001$ (worst-case family deviation 0.2 percentage points). On Llama-3.1-8B under GPT-4o judge ($n=100$ per bench, zero content-filter blocks), DIAL exceeds a FiLM baseline trained with the same multi-$c$ objective on every benchmark and judge. Semantic vs strict judge: HarmBench DIAL +0.52 / +0.55 vs FiLM +0.44 / +0.39; AdvBench DIAL +0.63 / +0.66 vs FiLM +0.46 / +0.50; XSTest DIAL +0.33 / +0.25 vs FiLM +0.26 / +0.22; both judges agree on the ranking. Outside training range, DIAL's residual saturates at $\sigma(g_{\max})$ while FiLM's grows unboundedly: at $c = 10^4$, FiLM's residual $L_\infty$ norm is 534 times larger than DIAL's. The bound holds across all control values, not only trained ones; concrete deployment scenarios (configuration error, multi-tenant policy ceilings, adversarial composition) are where this distinction matters in practice. The bounded-monotone form produces analogous saturation behavior on two continuous-control RL environments (HC-Vel $n=35$, Ant-Goal $n=30$) under matched oracle-$c$; the prior fails when $c$ must be inferred online (meta-RL). Ablations confirm both the bound and the saturating temperature are necessary; a frozen-gate variant retains the bound at lower controllability magnitude. We release adapter train reports and canonical bench generations, algebraic derivations (with a SymPy reproducibility script), the full GPT-4o judge cache (semantic and strict), and a human-labeled validation set ($n=50$, Cohen's kappa = 0.895). Full adapter checkpoints are deferred to camera-ready release.

Dynamic Reconfigurable Battery (DRB) systems enable flexible interconnection among cells through power electronic switches, crucial for electric vehicles and energy storage systems. However, sensor failures in practical deployments lead to partial observation loss, posing severe challenges for safe control. Existing methods assume full state observability and often adopt overly conservative strategies when observations are missing, failing to achieve effective trade-offs between safety and performance under uncertainty. More fundamentally, these methods lack mechanisms to adaptively adjust decision conservatism based on inference reliability, leading to unstable decisions under complex failure scenarios. We propose the $\textbf{D}$iffusion-$\textbf{G}$uided $\textbf{R}$isk-$\textbf{A}$daptive $\textbf{F}$ramework (DGRAF) to address this challenge. The framework handles spatiotemporally coupled failure patterns through joint diffusion inference with quality quantification, and integrates inference quality with distributional decision-making to enable automatic modulation between aggressive optimization and conservative maintenance. Through extensive simulations and hardware experiments, we demonstrate that DGRAF consistently outperforms baseline methods across complex failure patterns, achieving superior energy efficiency, decision stability, and risk resilience.

Scientific measurements inherently contain variance. Standard regression, however, treats noisy data as absolute ground truth, leading models to memorize noise rather than physical laws. This overfitting hampers generalization, particularly when transferring from abundant theoretical proxies to scarce experiments. We introduce DETS, a framework replacing rigid point estimation with Physical Tolerance Modeling. Our Interval-Censored Evidential Engine (ICEE) maximizes probability mass within acceptable error margins, explicitly decoupling aleatoric noise from epistemic uncertainty. Using this filtered uncertainty signal, a Thermodynamic Sampling strategy dynamically selects theoretical data to align source and target domains. Experiments across thermodynamics, drug affinity, and bandgap prediction show DETS outperforms state-of-the-art methods. Crucially, it exhibits superior robustness in high-noise, data-scarce settings.

Video multimodal large language models have made strong progress on open-ended video understanding, but they still lack precise local spatiotemporal perception. When two videos share almost the same global semantics and differ only in a short time span or a small region, current models often fail to find the change and provide reliable evidence. We propose **DeltaVid**, a verifiable proxy-task framework that enhances fine-grained spatiotemporal perception with cross-video differences. The key idea is to turn cross-video spot-the-difference into a trainable perception signal, where a model identifies local changes, judges temporal boundaries, and organizes spatial evidence by comparing similar videos. To make this signal scalable to train and reliable to evaluate, we further introduce DeltaVid-10K and DeltaVid-Bench, which convert controllable local differences in real videos into evidence-labeled training and test samples. Experiments show that DeltaVid substantially improves performance on cross-video difference understanding and transfers the learned local evidence ability to general video understanding benchmarks, including MMVU, MLVU, Video-MME, VideoHolmes, VideoMMMU, LVBench, TempCompass, and LongVideoBench. These results show that cross-video differences are not only an effective way to diagnose fine-grained perception failures, but also a scalable proxy supervision that moves Video MLLMs from coarse semantic understanding toward fine-grained spatiotemporal evidence reasoning.