NeurIPS 2026 Papers

We show that replacing the standard MSE denoising loss in diffusion models with a nonlinear transformation induced by an f-divergence yields a simple robust training surrogate that empirically improves performance under data contamination, with small additional computational overhead. The theoretical foundation rests on a local divergence construction: under the Gaussian reverse-kernel structure of DDPM, each per-step likelihood ratio follows a lognormal distribution parameterized by a scalar mismatch, so the conditional f-divergence at each step reduces to a one-dimensional function of the denoising error. Summing these local divergences yields a training objective that unifies diffusion training as divergence-induced weighted denoising, where the derivative of the induced divergence acts as a residual-space influence weight that controls the contribution of each sample. Bounded-influence divergences (Hellinger, negative exponential) suppress large-error samples, with Hellinger yielding an explicit exponential weight, connecting the framework to robust M-estimation. Empirically, on CIFAR-10 under 30% contamination, NED reduces FID from 93.0 (KL) to 77.5, while also outperforming standard robust losses such as Huber and clipped MSE.

Fine-tuning large language models (LLMs) on downstream tasks often weakens their previously aligned safety behavior, creating an inherent trade-off between task adaptation and safety preservation. Existing approaches attempt to mitigate this issue by identifying safety-relevant neurons, layers, or update directions in the original parameter space. However, safety-related information is often distributed across many directions in this space, causing downstream updates to overwrite parameters that encode safety behavior. We propose WSR-Tune, a Weight Space Reparameterization-based fine-tuning framework that preserves safety during downstream adaptation. Our approach constructs a safety-conditioned basis and reparameterizes the weight matrices such that safety-relevant information becomes concentrated in a small subset of directions. We then identify safety-critical directions and freeze them in the reparameterized space, while updating the remaining complementary directions for downstream adaptation. By explicitly structuring the parameter space to disentangle safety and task information, our method reduces interference between objectives and enables simultaneous preservation of safety and improvement in downstream performance. Experimental results show that our method achieves a more favorable trade-off between downstream performance and safety retention, demonstrating its effectiveness for reliable LLM fine-tuning.

Large language models trained with Reinforcement Learning (RL) with verifiable rewards exhibit strong reasoning ability and broad generalization, whereas models trained with Supervised Fine-Tuning (SFT) are often viewed as more prone to memorization and limited transfer. This paper rethinks this distinction through the lens of SFT training data. First, we study the role of data source and show that it is critical: a carefully mixed SFT dataset substantially outperforms data generated solely by a larger model. Second, we study the role of data scale and show that matching the number of correct rollouts between SFT and RL greatly improves SFT generalization, while matching the total rollout budget enables SFT to generalize as well as RL. Combining these two factors further enables SFT to generalize even better than RL. Third, by using LLM annotations to characterize the solution methods in training rollouts, we show that larger datasets cover more tail methods and that these tail methods provide generalizable reasoning signals. Finally, we support these empirical findings theoretically by analyzing the training dynamics of shallow transformers under both RL and SFT.

EEG foundation models (FMs) offer strong population-level priors but degrade on individual subjects due to anatomy-driven distribution shifts. We propose Rest-Tuning, a task-agnostic calibration framework that personalizes a population FM using just 1–3 minutes of unlabeled resting-state EEG. Rest-Tuning employs masked teacher-student alignment to produce a subject-calibrated backbone, reused as initialization across multiple tasks from the same subject. Across two multi-subject datasets and four EEG FMs, Rest-Tuning outperforms task-only adaptation, with up to 13.3% absolute gain in downstream task accuracy and over 40% average reduction in labeled task data. Calibration saturates quickly and requires under 3 minutes of compute per subject. Representation analyses show reduced subject-to-population distance and smoother LoRA adapter interpolation paths, indicating improved compatibility between rest- and task-adapted solutions. These results establish resting-state EEG as a practical, fast, and reusable calibration interface, enabling population EEG FMs to function as reliable, subject-specific BCIs and serve as reusable initialization for downstream task adaptation.

Onboard satellite models often require frequent updates, but the weights adapted to earlier data distributions can quickly become outdated. However, updating large-scale model parameters in orbit presents significant challenges due to the limited uplink bandwidth of Low Earth Orbit (LEO) satellite systems, particularly for hyperspectral satellite imagery, where high-dimensional spectral–spatial inputs lead to increased model size and update costs. Existing full fine-tuning methods are thus expensive to retrain and difficult to deploy under strict communication constraints. To address this challenge, we propose NE-LoRA, a parameter-efficient adaptation framework for bandwidth-constrained onboard hyperspectral model updates. NE-LoRA combines a primary low-rank branch with a nonlinear auxiliary branch to capture both global update trends and complex spectral–spatial variations. Additionally, we introduce a differentiated training strategy for multi-matrix adapters, motivated by the asymmetric initialization and gradient dynamics of different adapter matrices. Experiments on four hyperspectral datasets and three representative backbone models demonstrate that NE-LoRA consistently outperforms LoRA-based baselines and remains competitive with, and in several cases superior to, full fine-tuning. Across the evaluated settings, NE-LoRA updates only a small fraction of the total parameters on average while preserving low deployment overhead, offering a favorable accuracy–communication trade-off for onboard hyperspectral adaptation.

In recent years, large language models (LLMs) have demonstrated remarkable capability to generalize across diverse natural language processing tasks, inspiring the development of graph foundation models (GFMs) for large-scale pre-training. However, unlike language models with explicit token units, graphs lack a well-defined unit for generalization, making it challenging to design effective pre-training strategies. In this work, we propose REEF, a novel GFM framework that leverages relation tokens as the fundamental units. We construct a vocabulary of relation tokens to encode relational information within graphs. To accommodate diverse relations, we introduce two hypernetworks that adaptively generate the parameters of aggregators and classifiers in graph neural networks based on relation tokens. In addition, we design another hypernetwork to construct dataset-specific projectors and incorporate a dataset-level feature bias into the initial node representations, enhancing flexibility across different datasets with the same relation. Extensive experiments demonstrate that REEF consistently outperforms existing methods in both pre-training and transfer learning, highlighting its potential as a general-purpose graph foundation model. Our code is publicly available at https://anonymous.4open.science/r/REEF-16CF/.

Explicit reasoning models are trained to produce intermediate reasoning traces before final answers, but downstream fine-tuning is often performed on ordinary instruction--response data that contains no such traces. We show that this mismatch can induce *reasoning-trace collapse*: a fine-tuned model continues to produce plausible final answers while losing the structurally valid explicit reasoning traces that made it a reasoning model in the first place. We introduce a structural evaluation framework that separates answer correctness from reasoning-trace validity, measuring valid, empty, missing, and truncated reasoning alongside reasoning-conditioned task performance. Using this framework, we study open-weight reasoning models and find that standard supervised fine-tuning can rapidly suppress valid reasoning traces, and that answer-only metrics can substantially obscure this failure: in several settings, performance conditional on valid reasoning remains high while the rate of valid reasoning falls sharply. We further show that simple loss-masking strategies substantially mitigate collapse without requiring teacher-generated reasoning traces. These results suggest that evaluations of fine-tuned reasoning models should report structural reasoning reliability in addition to final-answer performance, especially when adaptation data does not contain explicit reasoning traces.

Recent 3D foundation models, such as DUSt3R, MASt3R, VGGT, $\pi^3$, and Depth Anything 3, provide strong feed-forward depth and pose estimates on pinhole imagery, but degrade sharply under fisheye camera geometry. We show that this failure is partly caused by a pinhole camera bias in the positional encodings of pretrained 3D foundation models, and propose RayTun3R, a lightweight camera adaptation approach. It keeps the pretrained network fixed and adapts only lightweight components tied to token position and camera geometry. RayTun3R learns parameter-efficient residual corrections to absolute and rotary positional encodings, together with parameter-free tokenization and corrections to prediction-grid coordinates that remove residual pinhole assumptions. The resulting adapter contains only 10,752 trainable parameters and can be learned from a short temporal segment using geometric losses. Once adapted, RayTun3R transfers effectively to the remaining frames of the sequence without incurring additional runtime costs. Across diverse fisheye datasets with fields of view from $110^\circ$ to $200^\circ$, our adapter reduces rotation error by $2-12\times$ relative to the unadapted model, outperforms LoRA while using $\sim14\times$ fewer trainable parameters, improves pose over adaptation-free baselines while avoiding their multi-view inference cost, and remains competitive on depth accuracy.

Time series foundation models (TSFMs) commonly adapt to new data by attaching a single trainable head to a frozen backbone, a one-size-fits-all setup that underfits heterogeneous regimes. Replacing the head with a mixture of experts is the standard upgrade, but on instance-normalized backbones (the dominant TSFM design class) it fails: routing entropy collapses to zero and one expert absorbs every input, a failure we call *normalization-induced routing collapse*. Standard MoE rescue mechanisms do not repair it, because the cause is in the router's input, not its optimization. Pre-encoder normalization strips the statistics a router would need to tell regimes apart. A mutual-information decomposition makes this precise and yields a signal-ratio that, computed before training, predicts dataset vulnerability (Spearman $\rho = -0.88$). Eight causal controls, including a vision-modality replication, isolate instance normalization as the cause. The prescription is a minimal causal intervention: *Raw-Routed Mixture of Adapters* (RR-MoA), which routes on the raw, pre-normalization input. Under a strictly frozen backbone, RR-MoA wins 54/54 comparisons against the strongest fixed adapter and significantly outperforms LoRA, TRACE, AdaMix, and full fine-tuning. The effect generalizes across six backbones and an imputation task. Frozen RR-MoA also beats full fine-tuning by 12–79% (the *Frozen Paradox*); two architecturally distinct variants confirm the principle generalizes beyond this specific router. Code is provided in the supplementary material.

Non-contrast CT (NCCT) is widely used in clinical practice, yet its limited soft-tissue contrast often renders visual information insufficient for accurate diagnosis. In routine radiology workflows, NCCT is commonly accompanied by contrast-enhanced CT, radiology reports, and pathology reports, which provide complementary knowledge at different levels, from structural and morpho-functional cues to semantic findings and pathological evidence. However, existing pretraining paradigms address this challenge only partially, each introducing complementary but still limited constraints on NCCT representations. To address this limitation, we propose RadOmni, a unified omni-pretraining framework that injects multimodal radiology knowledge into representation learning for non-contrast CT. RadOmni formulates a curriculum consolidation learning strategy, in which knowledge is progressively injected according to optimization difficulty and knowledge level through four stages: masked image modeling on non-contrast CT, cross-phase transfer of anatomical structures and enhanced patterns from contrast-enhanced CT, report-guided semantic learning, and pathology-informed diagnostic enhancement. To mitigate optimization conflicts across heterogeneous objectives and reduce knowledge forgetting, each stage retains and jointly optimizes the objectives from preceding stages, while early encoder layers are selectively frozen to preserve generic representations. To enable such multi-source pretraining, we curate Omni-CT10K, a large-scale CT dataset with more than 10K NCCT scans featuring clinically co-occurring multimodal supervision. Extensive experiments on Omni-CT10K, two in-house benchmarks from different centers, and the public MSD benchmark show that RadOmni consistently achieves strong performance across classification, segmentation, pathological TNM staging, and radiology report generation tasks. Further analyses demonstrate favorable scaling behavior, robust cross-center generalization, and transferable representations.

Large language models are increasingly deployed as long-lived agents that must adapt across users, tasks, domains, modalities, and feedback regimes without access to model weights. Existing black-box adaptation methods typically optimize a single prompt, maintain an undifferentiated memory, or rely on repeated rollout-heavy search. However, these designs struggle when streams of input are nonstationary, feedback is sparse, and failures from one task family can contaminate behavior on another. We introduce RIZZ (**R**outing **I**nteractions to Near **Z**ero-interference **Z**ones), a continual adaptation framework for compound language-model systems that learns entirely through verifier-gated memory, routing, and prompt compilation. RIZZ organizes input streams into dynamically spawned memory branches. At inference time, either while online or offline, a context-aware router selects or creates a branch that retrieves branch-local, global, graph-structured, and working-memory context, which is compiled into a bounded prompt together with retrieved task evidence. After the model acts, task verifiers score the output, and only verified interactions can update memory, promote reusable rules, demote harmful rules, or create anti-patterns. This yields a black-box agent that improves through persistent natural-language feedback while explicitly controlling interference. RIZZ targets the regime where adaptation must occur online under context budgets. Finally, we demonstrate the effectiveness of our framework against state-of-the-art baselines on competitive benchmarks.

Policies trained on offline logs can perform well in open-loop evaluation but fail in closed-loop deployment, because they never learn to recover from their own mistakes. Fine-tuning these policies typically requires either many human demonstrations or large-scale reinforcement learning rollouts, both of which are expensive. We propose RAHF, a two-stage closed-loop fine-tuning framework that reduces human effort by amplifying a small number of human corrections with a cheap verifiable reward. In Stage 1, a human operator monitors the policy during closed-loop rollout and intervenes at dangerous states, producing a small but targeted correction dataset that also keeps the training process safe. In Stage 2, the improved policy runs without any human involvement. At each step, the verifiable reward scores the policy's candidate proposals. We use these reward scores to construct group-normalized advantages and fine-tune the policy toward the higher-advantage proposals. To prevent the policy from forgetting its pretrained knowledge, we propose a Policy Adaptation Module (PAM) that freezes the pretrained weights and learns closed-loop corrections through a separate trainable branch. On BridgeSim-NavHard, RAHF with a human teacher reaches a driving score of 78.90, surpassing all baselines including DAgger, HG-DAgger, GRPO, and AWR. In addition, PAM reduces forgetting and enables zero-shot transfer to HUGSIM without any additional fine-tuning.

The development of tabular foundation models (TFMs) has accelerated in recent years, showing strong potential to outperform traditional ML methods for structured data. A key finding is that TFMs can be pretrained entirely on synthetic datasets, opening opportunities to design data generators that encourage desirable model properties. Prior work has mainly focused on crafting high-quality priors over generators to improve overall pretraining performance. Our insight is that parameterizing the generator distribution enables an adversarial, distributional robustness perspective: during training, we can adapt the generator to emphasize datasets that are particularly challenging for the model, which we formalize by introducing an optimality gap measure. Further, we develop a method to align the synthetic data generation with real-world datasets from a given domain, constraining the adversary to generate "realistic'" data. Together, these algorithms comprise the Robust and Domain-Adapted Tabular Foundation Models (RAD-TFM) pipeline, a model-agnostic adversarial training framework. Applied to the TabPFN V2 classifier, RAD-TFM improves performance across 6 diverse tabular benchmarks, with up to a 11\% increase in mean normalized AUC over the original TabPFN and other baseline algorithms, with only 100k additional training datasets, less than 0.1\% of the original pretraining data. These results highlight a promising new direction for targeted adversarial training and fine-tuning of TFMs using synthetic data alone.

Large language models (LLMs) have demonstrated remarkable performance in video question answering (VQA). To improve responsiveness and answer quality, modern LLM-based VQA services pre-extract multiple LLM-compatible modalities from videos before response generation. However, the performance impact of modality combinations (MCs) and extraction knob settings has barely been studied. Our analysis shows that both factors substantially affect answer quality and Time-to-First-Token (TTFT) latency. Furthermore, the best configuration depends on both video content and query semantics, necessitating adaptive strategies. We propose Prysma, the first modality adaptation framework for LLM-based VQA. Prysma intelligently adapts the modality configurations to optimize answer quality, subject to a Service Level Objective (SLO) on average TTFT. It combines an offline modality orchestrator and an online semantic adapter. The offline orchestrator constructs a video knowledge base to warm-start tuning, employs a latency-aware configuration pruner to reduce search space, and runs a two-stage tuner to optimize both video- and query-level configurations. The online adapter selects configurations in real time based on query semantics and tuning history. Extensive evaluations demonstrate that Prysma improves answer quality by up to 82% over SOTA methods under the same SLO setting.

Feed-forward 3D foundation models have recently reshaped multi-view reconstruction, yet dynamic 4D understanding still often relies on large-scale motion supervision or hand-crafted, architecture-specific heuristics. We ask whether frozen 3D foundation model tokens already encode motion-relevant information that can be exposed by lightweight readout functions, rather than re-learning motion-specific features with high-capacity decoders. Through probing on three representative backbones, we find that static--dynamic separation is decodable from frozen token representations. Our rank-restricted and spectral analyses on $\pi^3$X reveal that the recovered motion cues are concentrated in a low-dimensional subspace. Building on this, we propose Probing-based Motion Distillation (PMD), which converts pairwise motion-state compatibility scores against a fixed anchor token into dense motion masks. Trained on only 160 YouTube-VOS videos using a single 16 GB GPU, PMD transfers zero-shot to DAVIS, SegTrackv2, and FBMS-59 without target-benchmark fine-tuning. With multi-view pretrained geometric backbones, PMD outperforms training-free heuristics and approaches substantially heavier supervised and self-supervised methods, suggesting that motion decoding from frozen 3D foundation models can be lightweight and data-efficient.

Large-scale pretraining has made Vision-Language-Action (VLA) models promising foundations for generalist robot manipulation, yet adapting them to downstream tasks remains necessary. However, the common practice of full fine-tuning treats pretraining as initialization and can shift broad priors toward narrow training-distribution patterns. We propose PriorVLA, a novel framework that preserves pretrained priors and learns to leverage them for effective adaptation. PriorVLA keeps a frozen Prior Expert as a read-only prior source and trains an Adaptation Expert for downstream specialization. Expert Queries capture scene priors from the pretrained VLM and motor priors from the Prior Expert, integrating both into the Adaptation Expert to guide adaptation. Together, PriorVLA updates only 25% of the parameters updated by full fine-tuning. Across RoboTwin 2.0, LIBERO, and real-world tasks, PriorVLA achieves stronger overall performance than full fine-tuning and state-of-the-art VLA baselines, with the largest gains under out-of-distribution (OOD) and few-shot settings. PriorVLA improves over $\pi_{0.5}$ by 11 points on RoboTwin 2.0-Hard and achieves 99.1% average success on LIBERO. Across eight real-world tasks and two embodiments, PriorVLA reaches 81% in-distribution (ID) and 57% OOD success with standard data. With only 10 demonstrations per task, PriorVLA reaches 48% ID and 32% OOD success, surpassing $\pi_{0.5}$ by 24 and 22 points, respectively.

Supervised fine-tuning (SFT) is computationally efficient and broadly applicable, but often shows weaker generalization than reinforcement learning (RL). A key limitation is its off-policy objective: SFT fits fixed demonstrations token by token, including targets that may be poorly aligned with the model's pretrained distribution. Recent token-reweighted SFT methods address this issue by assigning larger training weights to tokens that better align with the model's predictive distribution, using statistics such as target-token probability or entropy. However, computing these statistics from the online model being fine-tuned makes token weights trajectory-dependent, as the model's distribution rapidly departs from the pretrained model and induces self-reinforcing reweighting dynamics. We propose PriFT, Prior-support guided Fine-Tuning, which derives token weights from a frozen pretrained reference to obtain a stable reweighting signal unaffected by online fine-tuning dynamics. This signal estimates prior support: the extent to which each target token is supported by the pretrained model before task-specific adaptation. Across multiple existing weighting and selection rules, replacing online statistics with pretrained statistics consistently improves performance. We introduce two instantiations: PriFT-prob, which uses pretrained target-token probability, and PriFT-mass, which selects tokens by relative support under the pretrained distribution. Extensive experiments on mathematical reasoning, code generation, and medical question answering show that PriFT achieves state-of-the-art SFT results and provides a better initialization for subsequent RL training.

Transformers follow _implicit_ curricula whereby some tasks are learned before others. However, how _explicit_ pretraining curricula influence learning, generalization, and the selectivity of fine-tuning is unclear. This is important for AI safety, where fine-tuning is used to selectively suppress misaligned behaviors. Here, we compare curricula that pretrain tasks in a balanced (sampled uniformly) or an imbalanced (one task early, the other late) fashion. We show that imbalanced learning of two conflicting copy tasks promotes in-context learning and improves the selectivity of refusal fine-tuning. Ablations and activation patching show that this occurs because imbalanced pretraining encourages tasks to be disentangled in separable neural circuits, whereas balanced training routes both tasks through a common pathway. We extend these findings to a synthetic language learning task involving rule-consistent and rule-violating data, where imbalanced curricula similarly lead to more localized, less entangled rule representations, resulting in more robust rule-following behavior. Together, these results suggest that imbalanced pretraining curricula may be an important tool for promoting disentangled representations, with direct consequences for the precision and reliability of safety fine-tuning.

Neuro-symbolic learning was proposed to address challenges with training neural networks for complex reasoning tasks with the added benefits of interpretability, reliability, and efficiency. Neuro-symbolic learning methods traditionally train neural models in conjunction with symbolic programs, but they face significant challenges that limit them to simplistic problems. On the other hand, purely-neural foundation models now reach state-of-the-art performance through prompting rather than training, but they are often unreliable and lack interpretability. Supplementing foundation models with symbolic programs, which we call neuro-symbolic prompting, provides a way to use these models for complex reasoning tasks. Doing so raises the question: What role does specialized model training as part of neuro-symbolic learning have in the age of foundation models? To explore this question, we highlight three pitfalls of traditional neuro-symbolic learning with respect to the compute, data, and programs leading to generalization problems. This position paper argues that foundation models enable generalizable neuro-symbolic solutions, offering a path towards achieving the original goals of neuro-symbolic learning without the downsides of training from scratch.