NeurIPS 2026 Papers

Recent think-answer approaches in VLMs, such as Qwen3-VL-Thinking, boost reasoning performance by leveraging intermediate thinking steps before the final answer, but their high computational cost limits real-world deployment. To distill such capabilities into compact think-answer VLMs, a primary objective is to improve the student's ability to utilize visual evidence throughout its reasoning trace. To this end, we introduce a novel think-answer distillation framework that encourages the student to anchor its thinking on visual information by masking the student's salient reasoning prefixes. To compensate for such masked textual cues, the student is encouraged to rely more on visual evidence as an alternative source of information during distillation. Our masking strategies include: 1) $\textit{token-wise salient reasoning-prefix masking}$, which masks high-influence reasoning prefixes selectively for each next-token prediction, and 2) $\textit{self-paced masking budget scheduling}$, which gradually increases the masking scale according to distillation difficulty, {measured by discrepancy between teacher--student distributions. In the distillation phase, the student is guided by our salient reasoning-prefix mask, which blocks both future tokens and salient reasoning cues, in place of the standard causal mask used for auto-regressive language modeling. Experimental results show that our approach outperforms recent open-source VLMs, VLM distillation, and self-distillation methods on multimodal reasoning benchmarks, while further analyses confirm enhanced visual utilization along the student thinking process.

Reinforcement learning with verifiable rewards (RLVR) has become a powerful paradigm for improving language models on reasoning-intensive tasks, but its effectiveness is often limited by exploration. For example, models often fail on hard problems, leaving little useful reward signal. External expert traces offer a natural source of guidance, yet they may also expose reward-relevant content along the critical path to the verifier target, such as final answers, intermediate values, executable implementations, or answer-related entities. This content can create an unintended reward hacking channel, allowing the policy to obtain reward by copying the trace rather than learning the underlying reasoning or agentic behavior. Existing guided-RL methods reduce this risk by using partial trajectories, but they mainly control how much expert information is shown heuristically rather than which parts should be hidden. To this end, we propose Semantic Masked Expert Policy Optimization (SMEPO), a fine-grained semantic masking strategy for expert-guided RLVR. Instead of truncating traces coarsely or revealing them unchanged, SMEPO masks reward-relevant semantic spans along the critical path while preserving the expert’s decomposition, plan, and procedural structure. This turns hard problems from reasoning from scratch into a fill-in-the-blank process: the policy can follow the expert’s problem-solving route, but must still reconstruct the missing values, code, or entities by itself. SMEPO is simple to apply and requires no changes to the reward function or RL objective. Across diverse domains, including math, code, and agentic search, SMEPO improves accuracy by up to 3.2 points over GRPO and reduces training time by up to 4.2$\times$.

Knowledge graph question answering (KGQA) systems are typically evaluated on questions whose surface tokens lexically expose the gold KG path. Real users, however, ask questions without knowing the underlying schema: "Which condition could Andrew have due to family history?" compresses a 4-hop chain (Andrew - father → father → medical condition → subclass of - hereditary disorder) into a single phrase (family history) whose tokens name none of those relations. We show that this collapse of path-question alignment is a blind spot of existing KGQA evaluation. The three dominant Large Language Model (LLM)-KG paradigms — hop-by-hop KG exploration, plan-then-retrieve, and similarity-based retrieval — all implicitly rely on the question's surface leaking the gold path, an assumption rarely tested in practice. We make three contributions. (i) HiddenPathQA, a human-validated KGQA benchmark of 1,055 multi-hop questions (2-6 hops, multiple domains) constructed so that the question surface does not leak the gold KG path's relations. (ii) Holistic Path Selection, a reasoning paradigm that exposes all candidate relation chains around the topic entity to the LLM at once, replacing per-hop relation scoring with question-against-full-chain alignment. Two implementations, TieredHolistic and FlatHolistic, outperform six strong baselines (ToG, PoG, FiDeLiS, R2-KG, KAPING, KARPA) on HiddenPathQA by +8.8 to +17.9 points over the strongest baseline across all four (backbone × split) settings. (iii) A pseudoword stress test that replaces entity surfaces with placeholder tokens. It both confirms HiddenPathQA items are solvable from KG structure alone — not from entity-level priors — and reveals that several baselines rely on entity memorization rather than KG-grounded reasoning.

High-stakes prediction systems are often evaluated only on labels revealed by past decisions. This selective observation can hide the losses that determine full-population tail risk: a predictor may appear safe on revealed labels while its worst compatible failures remain unseen. We study certification of conditional value-at-risk (CVaR) under selective labels. First, we prove non-identifiability: even with overlap, two predictors with identical observed selected-loss distributions can have reversed full-population CVaR rankings under compatible hidden-label laws. Under an outcome-dependent odds-ratio sensitivity model, we derive sharp upper and lower CVaR envelopes. A minimax interchange shows that worst-compatible hidden-label completion commutes with the Rockafellar--Uryasev threshold optimization, giving an exact certificate rather than a loose robust surrogate. We then define the tail-risk certification frontier: the minimum passive revelation cost needed to reduce hidden-tail ambiguity below a target tolerance. For a fixed predictor and threshold, this frontier is a fractional-knapsack problem whose value density, tail-identification value, pinpoints labels that can move the CVaR tail. Finally, we give finite-sample guarantees for cross-fitted conservative certificate learning and a lower bound governed by the effective number of revealed tail labels.

Text-code embeddings underpin modern code intelligence systems, but progress in this area is held back by a systemic problem: code retrieval benchmarks routinely treat valid alternative implementations as negatives, making cross-model comparison unstable and depressing measured performance. We document this problem with a detailed analysis of six widely used benchmarks and show that, depending on the dataset, between 44\% and 87\% of items a strong LLM judge marks as relevant are missing from the original labels. To address this, we adopt a unified evaluation protocol in which a single LLM judge re-annotates the top-10 retrieved candidates for every model under a fixed prompt, replacing fragmented benchmark heuristics with a consistent semantic standard. Applying the same idea at training time, we use a smaller LLM-supervised reranker to construct MegaCode, the largest semantically curated dataset for code search to date of 193M positive (text, code) pairs, roughly an order of magnitude larger than prior curated corpora, and train a family of embedding models on it at four scales (0.5B-7B) with contrastive learning. Under a unified LLM-judged protocol, our 0.5B model matches or exceeds all prior open code-retrieval embeddings up to 8B parameters, and our 7B model improves average MRR@10 by 3.5 points over the strongest baseline.

LLM evaluations drive which models get deployed, what safety standards get adopted, which research conclusions get published, and how projections of AI's labor-market impact get made. Yet standard confidence intervals ignore variability from judge model choice, model temperature, and prompt phrasing, producing under-coverage that worsens with more data. The omitted variance can shift results enough to reverse conclusions \citep{baumann2025llmhacking, huang2026dropping}; pipelines that fail to average over it leave the surface that ``benchmark hacking'' exploits \citep{singh2025leaderboard}. This paper decomposes LLM pipeline uncertainty into its sources, distinguishes variance that shrinks with more data from sensitivity to researcher design choices, and uses design-study projections to reduce total evaluation error (TEE). Across the demonstrations, naive standard errors are 40 - 60\% smaller than the TEE-corrected SE. Using Chatbot Arena data, we show naive 95\% CI coverage drops as $n$ grows while TEE-corrected coverage holds at 95\%, and TEE-guided pipelines restrict the benchmark gaming surface from 56 to 32 Elo ($K=27$), below the human-leaderboard baseline. We show further that a small pilot recovers honest CIs and projects which design changes most improve precision. Acting on those projections halves MMLU estimation error against the answer key at equivalent cost, and raises per-match agreement with human votes by 7.9 percentage points on Chatbot Arena.

Multimodal large language models must continually adapt to evolving tasks and domains, yet standard continual learning metrics mainly measure whether old answers remain correct, leaving the stability of multimodal grounding largely unexamined. We study this overlooked failure mode and ask whether a continually adapted MLLM can preserve not only what it answers, but also how it uses visual, textual, OCR, chart, and document evidence. We identify hidden evidence-use forgetting, where answer accuracy is retained while the model silently shifts toward different or less grounded evidence channels, and propose RCL, a replay-free reliance-constrained continual learning framework. RCL freezes the previous checkpoint as a behavioral reference, estimates teacher and student evidence-reliance profiles through counterfactual channel interventions, and jointly optimizes task learning, prediction preservation, and reliance preservation without adding inference-time cost. Across CoIN, COAST, MCITlib, and an evidence-sensitive multimodal stream, RCL consistently improves final performance and reduces forgetting over replay-free, PEFT, routing, and memory-assisted baselines, while substantially lowering modality reliance drift, dominant evidence flips, and hidden forgetting rates. These results suggest that robust continual multimodal learning requires preserving the evidence path behind correct answers, not merely the answers themselves.

Video generation models have achieved remarkable visual fidelity and have strong potential to become general-purpose world simulators. Despite this progress, they still fail to generate videos which adhere to laws of physics. The problem becomes even more apparent in realistic settings where multiple physical principles must work together within the same video; for example, "a balloon floating upward while steam rises from a pot" requires buoyancy and fluid dynamics to unfold coherently and simultaneously. Yet existing methods largely ignore multi-principle interactions, focusing on a single principle per video. We propose HiPhy (Hierarchical Physical Alignment), a reinforcement learning framework that grounds video generation in physical laws through a dual-level objective: locally enforcing the temporal dynamics of individual physical principles, and globally ensuring the physical and semantic coherence of the entire scene. To support multi-principle generation, we construct a 50K-prompt dataset and introduce a prompt benchmark MultiPhyBench, spanning a diverse range of co-occurring physical events. Our experiments show that HiPhy significantly outperforms prior methods and baselines, improving physical commonsense and semantic alignment significantly across various benchmarks, with the largest gains on scenes involving multiple concurrent physical principles where competing methods degrade most sharply.

Existing multimodal large language models (MLLMs) typically formulate object detection as a conditional sequence generation task. Under this paradigm, spatial locations are commonly represented by either textual coordinates or special quantized learnable tokens. However, textual coordinates are tokenized as unrelated symbols, spatially adjacent points (e.g., $x=199$ vs. $y=200$) yield entirely different tokens, so the cross-entropy loss is misaligned with true spatial distance. Special quantized location tokens, on the other hand, inevitably introduce quantization error. Reducing this error requires linearly more tokens, but a larger vocabulary is harder to train and converges slower. We argue that both limitations stem from the formulation of spatial representation. To address this issue, we propose a simple yet novel localization representation method, HiLoc, which predicts locations in a coarse-to-fine manner with special tokens: first a grid index for coarse localization, then a cell index to refine the position within the grid. Under this paradigm, accuracy can be scaled up by simply adding the hierarchy level. We also build HiRef-500K, a dataset consisting negative and complex referring samples to strengthen rejection and complex referring ability. Experiments show that our method achieves strong performance on both detection and grounding benchmarks, especially on small objects, under high IOU thresholds and complex referring tasks.

Low-Rank Adaptation (LoRA) has emerged as a widely used technique for adapting large language models (LLMs) to new domains, due to its modular design and broad availability on platforms such as HuggingFace. This availability has motivated efforts to reuse existing LoRAs for domain generalization. However, existing methods often rely on explicit task labels or additional training, which are impractical for deployment. Moreover, they typically activate a fixed number of entire LoRA, leading to parameter redundancy or insufficiency that degrade performance. In this paper, we propose HiLoRA, a training-free framework that performs adaptive hierarchical routing over LoRA pools. Drawing on structural properties of LoRA, we define rank-one components (ROCs), where each LoRA rank is treated as an independent unit. For a given input, HiLoRA first adaptively selects a subset of LoRAs and determines their ROC allocation based on Gaussian likelihoods at the sequence level. At the token level, it refines routing by activating only the most informative ROCs. We further provide theoretical guarantees that HiLoRA selects the most relevant LoRAs with high probability. Extensive experiments show that HiLoRA achieves substantial improvements in domain generalization, with accuracy gains of up to 70\% over state-of-the-art baselines, while maintaining comparable inference throughput.

The advancement of graph-based malware analysis is critically limited by the absence of large-scale datasets that capture the inherent hierarchical structure of software. Existing methods often oversimplify programs into single-level graphs, failing to model the crucial semantic relationship between high-level functional interactions and low-level instruction logic. To bridge this gap, we introduce HiGraph, the largest public hierarchical graph dataset for malware analysis, comprising over 200M Control Flow Graphs (CFGs) nested within 499K Function Call Graphs (FCGs). This two-level representation preserves structural semantics essential for building robust detectors resilient to code obfuscation and malware evolution. We demonstrate HiGraph’s utility through a large-scale analysis that reveals distinct structural properties of benign and malicious software, establishing it as a foundational benchmark for the community. The dataset and tools are publicly available at https://higraph.org.

Training large foundation models at low numerical precision is one of the most promising directions for reducing the compute and memory cost of modern AI. Recent 4-bit floating-point formats such as MXFP4 and NVFP4 can be applied to linear GEMM operations in LLMs, but their limited dynamic range introduces numerical instability that prior work addresses by stacking stabilization mechanisms, typically executed at higher precision and partially eroding the efficiency gains that motivate FP4. In this work, we argue that numerical format design is itself a first-class lever for stable FP4 training, and present the first systematic study of FP4 LLM pretraining on energy-efficient Huawei Ascend NPUs. We compare the recently proposed HiFloat4 (HiF4) format against MXFP4 across both dense (OpenPangu-1B, Llama3-8B) and Mixture-of-Experts (Qwen3-MoE-30B) architectures, executing all linear and expert GEMMs in FP4. HiF4's hierarchical scaling provides enough representational headroom that a single stabilization step suffices to keep the relative loss within 1\% of a full-precision baseline — roughly half the gap of MXFP4, which requires three such mechanisms — while incurring less than 1\% average degradation on downstream tasks. Our results suggest that stable, accurate FP4 training does not require an ever-growing stack of stabilization techniques; it requires the right numerical format.

We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision. A systematic study reveals that the dominant source of degradation in FP4 RL is not training-side quantization error but rollout activation quantization: outliers stretch the dynamic range so far that a large number of activation values underflow to zero under FP4. Counterintuitively, restoring the training policy to higher precision while keeping the rollout in FP4 makes accuracy worse than full FP4 baseline, exposing rollout–training mismatch as the principal failure mode and ruling out standard pretraining-style fixes. We address this with Rollout Residual Quantization (Rollout-ResQ): a single residual correction term constrained to a hardware-friendly sparsity pattern, added only to the FP4 rollout matmul — a lightweight correction that recovers most of the precision lost to outlier-driven underflow without inflating the rollout’s compute footprint. On Qwen2.5-3B and Qwen2.5-Math-7B, Rollout-ResQ paired with the HiFloat4 (HiF4) format — whose three-level hierarchical scaling preserves resolution under FP4’s tight 4-bit budget — closes the accuracy gap to BF16 from 4.9% to 1.1%, bringing fully quantized FP4 RL within striking distance of full precision. Applied to the open-standard MXFP4, the same recipe narrows the gap from 13.6% to 5.3%, revealing that FP4 format choice is a key factor that determines the ceiling on recoverable accuracy. Together, these results establish HiF4 as the enabling format for end-to- end FP4 RL post-training, and Rollout-ResQ as the activation-side mechanism that makes the gap to BF16 closable.

Autonomous AI research promises to accelerate the scientific progress of machine learning. While current Large Language Model (LLM)-based agents excel at writing code, the bottleneck in research is the exploration of diverse and novel ideas: agents collapse to common techniques present in their pretraining data, or prematurely converge on suboptimal solutions (Padmakumar et al., 2024; Jiang et al., 2025). To this end, we introduce Heuresis, a framework that abstracts the research pipeline into a set of general and composable primitives, enabling open-ended scientific exploration in machine learning research. We implement 5 known algorithms spanning Quality-Diversity, Evolutionary, and Curiosity-based search, in addition to a greedy baseline, and evaluate them across three axes -- Quality, Diversity, and Novelty -- on two domains: LLM pretraining and On-Policy RL. We find that quality and novelty are inversely correlated. Methods that optimize purely for raw performance reach the strongest single solutions on a given task by replicating prior work: the top-quality runs from the greedy baseline are uniformly classified as direct copies under the Gupta-Pruthi rubric (Gupta & Pruthi, 2025). The few high-quality, verified-novel ideas in our study come from algorithms that balance performance with diversity or curiosity-based exploration. We also observed that agents resorted to a variety of reward-hacking techniques during execution, whose detection was necessary to keep the search faithful to the task. Our results underscore the importance of search and Quality-Diversity methods for autonomous research, and our framework opens up opportunities for further inquiry towards the ultimate goal of perpetual, autonomous scientific progress.

Foundation model training is becoming multimodal across the stack, from natively multimodal post-training pipelines to large-scale pretraining. As multimodal coverage broadens, context windows grow, and encoder–LLM scales diverge, a single LLM-centric TP/DP/PP/CP layout increasingly bottlenecks training throughput. This coupling forces encoders to inherit LLM-driven sharding and placement choices that can introduce unnecessary communication, limit useful encoder parallelism, or constrain the LLM schedule; the mismatch is especially pronounced at long contexts, where LLM context parallelism is needed for the fused multimodal sequence but encoder inputs remain bounded. We present heterogeneous parallelism for multimodal large language model training, a training-system abstraction that lets modules in the same end-to-end graph use independent parallel layouts and rank placements, supporting colocated execution, where modules share physical GPUs under different logical grids, and non-colocated execution, where modules occupy disjoint rank sets. The key systems challenge is preserving boundary tensor semantics when adjacent modules use independent layouts: forward activations must be materialized for the destination layout, while backward gradients must be routed back to the source layout. We address this with boundary communicators, runtime primitives that implement these forward and backward layout transforms, together with scheduling extensions for both placement modes. We evaluate optimized homogeneous, colocated heterogeneous, and non-colocated heterogeneous configurations across diverse multimodal workloads and GPU scales to characterize where each placement mode helps. Across this sweep, colocated heterogeneity improves TFLOPs/s/GPU by up to 14.4\% at short context and 41.8\% at long context, while non-colocated heterogeneity improves aggregate token throughput by up to 13.0\% and TFLOPs/s/GPU by up to 9.6\%. We validate loss convergence parity against homogeneous baselines and release an open-source implementation in Megatron-LM.

Pairwise comparisons from multiple judges are central to large language model evaluation and preference modeling, yet standard ranking pipelines often pool judgments into a single score vector, treating systematic judge disagreement as noise. We propose Heterogeneous Judge-Aware (HJA) ranking, a structured multi-judge ranking framework that separates consensus ranking, judge-specific sensitivity to consensus, and residual preference disagreement. HJA thereby treats ranking, judge sensitivity, and structured disagreement as separate inferential targets. We establish conditions under which this decomposition is identifiable and develop an anchored alternating algorithm that preserves the identifying geometry. For confidence quantification, we study a fixed-panel repeated-comparison regime in which the judge panel may remain fixed or modest while information grows through repeated judgments. This yields uncertainty statements for consensus and judge-specific ranking contrasts, sensitivity parameters, pairwise probabilities, and summaries of residual disagreement.Experiments on synthetic and real multi-judge comparison data show that HJA improves recovery, robustness, uncertainty calibration, and near-tie performance relative to pooled and sensitivity-only baselines. The fitted model also provides diagnostics for judge disagreement and model-affinity patterns, giving a statistically grounded framework for ranking under heterogeneous comparative judgments.

We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new Reinforcement Learning from Verifiable Reward (RLVR) problem that addresses the inefficiencies of isolated multi-agent on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional homogeneous teacher-to-student transfer. Building on this problem, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO with double rollouts by an average of 3.6% while using only half the rollout cost.

Federated Continual Learning (FCL) requires distributed clients to acquire new tasks while retaining previously learned knowledge. In this paper, we revisit the heterogeneity challenge in FCL, which is composed of two distinct aspects: spatial static heterogeneity, e.g., initial data distribution and model architecture differences, and temporal dynamic heterogeneity, e.g., task shifts over time. While existing methods acknowledge the presence of heterogeneity, they typically treat it in a unified manner by employing a single adaptive strategy without explicitly disentangling their impacts. In contrast, we propose a heterogeneity-aware distillation framework for FCL called Ha-FCL, which is designed to handle spatial and temporal heterogeneity through tailored distillation strategies. Specifically, Ha-FCL combines width-aware federated training with server-side public-anchor distillation, integrating multi-layer feature distillation for architecture-aware alignment with historical-logit regularization and representation co-distillation. This forms a coordinated dual-dimensional distillation scheme that balances adaptation to new tasks with retention of prior knowledge. This enables more targeted knowledge transfer and leads to improved model performance under complex FCL scenarios. We evaluate our method in various heterogeneous settings. Compared with state-of-the-art baselines, Ha-FCL consistently improves test accuracy and shows better robustness under challenging heterogeneous FCL settings.

Many real-world heterogeneous graphs exhibit pronounced heterophily, where connected nodes often have dissimilar labels or play different semantic roles. In such settings, standard heterogeneous graph neural networks that aggregate messages along metapaths or meta-relations primarily based on feature similarity can propagate misleading information, since feature similarity may be misaligned with underlying relational semantics. In this paper, we propose **HeterSEED**, a semantics–structure decoupling framework for heterogeneous graph learning under heterophily. HeterSEED decouples representation learning into a heterogeneous semantic channel that captures type- and relation-aware local semantics and a structure-aware heterophily channel that separates homophilic and heterophilic neighborhoods via pseudo-label-guided partitioning and aggregates them using metapath-based structural weights. A node-level adaptive fusion mechanism then combines the two channels to produce context-dependent node representations. Theoretically, we establish that, on heterogeneous graphs under heterophily, HeterSEED is strictly more expressive than standard heterogeneous graph neural networks that rely primarily on feature similarity and provably reduces the prediction bias introduced by heterophilic neighbors. Experiments on five real-world heterogeneous graphs, including two large-scale networks at the million-node and hundred-million-edge scale, demonstrate that HeterSEED consistently outperforms representative heterogeneous graph neural networks and recent heterophily-aware baselines, especially in strongly heterophilic regimes.