NeurIPS 2026 论文列表

Fine-tuning large transformers routes gradients through every token position, paying the cost in activation memory that scales linearly with sequence length and dominates the GPU budget. Existing efficient methods reduce the parameter side---low-rank adapters, prefix tuning, weight-decomposed updates---but leave the activation footprint untouched. We present the Token Hypothesis, a memory-efficient fine-tuning procedure motivated by a Neural Tangent Kernel analysis of how fine-tuning signal is distributed across token positions. We prove that every position is a low-dimensional capacity bottleneck whose width does not grow with the model, that positions with aligned hidden-state geometry are interchangeable in bidirectional models, and that causal positions form a monotone capacity chain in which later positions strictly dominate earlier ones. The analysis pins down how many positions are needed and which ones, both computable from forward passes alone: any subset suffices in bidirectional models, while the last few positions suffice in causal ones. We translate this into a procedure that backpropagates through only a small subset of positions while running the full forward pass over the rest. Empirically, the procedure matches or exceeds full-sequence fine-tuning across six task families---commonsense reasoning, mathematical reasoning, visual question answering, video question answering, GLUE, and image classification---and five backbones spanning bidirectional and causal architectures. It reduces activation memory substantially, translating to up to ∼ 20% lower total GPU memory (∼ 33GB) on long-context tasks, and composes cleanly with parameter-efficient methods such as LoRA and DoRA.

Parameter-efficient fine-tuning (PEFT) and probing enable adaptation of foundation models using only a small number of trainable parameters, making it attractive for video understanding where annotation and computation are expensive. However, video PEFT has focused on adapting image-pretrained models, while standard PEFT methods can also be applied to video representations. These settings are rarely compared and both confine temporal reasoning to a single component of the model, leaving open how temporal context should be distributed across backbone, PEFT and probe. In this work we provide a systematic study of model adaptation strategies for video understanding. We evaluate methods across appearance-focused, motion-focused and spatially dense settings, with a particular focus on scenarios with limited data where parameter-efficiency is most beneficial. Our results provide new insights into PEFT and probing across settings and demonstrate the importance of temporal context allocation for effective video adaptation.

Tokenizer transplant in cross-vocabulary model composition reconstructs donor-only embedding rows as weighted combinations over shared lexical anchors and reuses those coefficients on the base. We identify a structural geometric property of this reconstruction: the same coefficient vector reaches different sets in the donor and base anchor spans, an asymmetric realizability gap. Across 65 donor-base pairs under OMP, with cross-operator validation on CLP, WECHSEL, and FOCUS, we construct breaker tokens: single coefficient vectors that remain statistically inert in the donor anchor span while producing a high-salience reconstruction in the base. The same Gemma-2-2B donor checkpoint admits this construction against 13 different downstream bases drawn from five model families. The planted direction passes weight-merging with a clean reference unchanged. In a deployer case study, standard LoRA fine-tuning suppresses the breaker primarily on prompts whose distribution matches the training corpus and is not a sufficient mitigation against this attack family in our setting. The tested spectral filters miss the asymmetry. We discuss potential misuse in the open-weight composition supply chain.

Rehearsal-free continual learning with parameter-efficient adapters can be cast as a sequence of task-vector write-in operations: for each new task, a low-rank adapter is learned and merged into a running model. We propose Proximity Regularized Merging (PRM), a minimal modification to sequential LoRA merging that adds a proximal penalty during task-vector training without changing the subsequent write-in rule. PRM acts as a robust task-vector regularizer: in the reported Base$\rightarrow$+Prox diagnostics, it improves AAA across multiple write-in rules, backbones, and class-incremental settings, while its fixed-coefficient variant remains competitive with strong coefficient-based baselines. Mechanistically, matched-prefix norm controls and proximal-strength sweeps show that proximal training shrinks the task-vector radius, lowers Fisher-weighted interference, broadens the coefficient plateau, and exposes a stability--plasticity trade-off. Together, these results suggest that the effectiveness of sequential LoRA merging depends not only on how much of a task vector is written in, but also on whether the task vector itself has been trained to be mergeable.

Low-Rank Adaptation (LoRA) constrains weight updates to a low-rank subspace, making the choice of that subspace central to adaptation. We show that the value of LoRA initialization direction is regime-dependent. It helps when the representation dimension is large relative to the labeled set size and the estimated activation subspace is aligned with task structure, and is otherwise inconsequential. We formalize this view with a spiked-covariance Davis--Kahan argument and a first-step LoRA analysis showing that the initial learning signal flows through projected target-token features. From this analysis we derive an alignment coefficient $\rho$, computable from a single forward pass, that closely tracks the empirical gain from direction-aware initialization across tasks. Motivated by this view, we propose PivotLoRA, an initialization that estimates LoRA adapters from target-token PCA and applies a single output-preserving orthogonal pivot after early training drift. Across $30$ BERT-base few-shot settings and $13$ compared methods, PivotLoRA achieves the best average rank and improves over vanilla LoRA on nearly every setting, with the largest gains concentrated on high-alignment tasks. Decoder-scale classification on LLaMA-2-7B and Llama-3.1-8B reproduces the low-shot pattern, while full-data and generation controls confirm the predicted regime specificity.

Large language models may encounter factual knowledge during pre-training yet fail to reliably use that knowledge after fine-tuning. We study this gap in a stylized one-layer self-attention + MLP transformer trained by next-token prediction and subsequently fine-tuned on question-answering data. We first prove that, under suitable regularity conditions, the model reaches near-optimal pre-training loss while learning structured attention patterns. We further show that fine-tuning turns the Q\&A prompt format into a trigger for pre-trained relation features, enabling the model to extract facts not revisited during fine-tuning. Our analysis reveals a relation-covering characterization for knowledge extraction: fine-tuning need not revisit every stored fact, but it must cover enough latent relation-template directions through which facts were encoded during pre-training. We prove that extraction improves with pre-training multiplicity and fine-tuning coverage, but becomes harder as the relation-template universe grows. Conversely, insufficient coverage yields a failure regime in which facts can be stored but not extracted, providing a stylized mechanism for hallucination. Our analysis covers both full and low-rank fine-tuning, and experiments on synthetic data and PopQA-based GPT-2/Llama models support the predicted trends.

Semivalues are widely used to assign credit and guide data curation decisions, yet their outputs depend on a utility function that is not uniquely determined by the task. We study partition-based decisions, which directly model data curation tasks such as selecting high-quality subsets or flagging noisy examples, under two structurally unavoidable sources of utility underspecification: monotone transformations of performance scores and unconstrained small-sample behavior. We formalize outcome robustness in this setting as invariance of the induced partition across admissible utility respecifications. We establish that partition robustness is necessary for guaranteed success of semivalue-based data selection. To assess this condition in practice, we provide algorithms that certify partition robustness or produce a concrete witness utility demonstrating instability, requiring no utility evaluations beyond those already computed for the semivalues themselves, with formal correctness guarantees under both exact and approximate computation. We show that utilities admitting a coalitionally dominant top-$k$ subset, where high-quality points contribute more than low-quality points across every possible coalition, guarantee robust partition decisions, subsuming the modular utility case previously identified as sufficient. Together, these results provide a practical framework for determining when semivalue-based selection is well-defined under utility ambiguity.

Vision-Language Models (VLMs) such as CLIP have become a standard backbone for open-vocabulary recognition, yet their zero-shot predictions remain vulnerable to distribution shifts encountered at deployment. Test-Time Adaptation (TTA) has recently been extended to CLIP as a lightweight solution, leading to a rapidly growing body of TTA4CLIP methods. However, empirical progress in this area has largely outpaced our understanding of what truly drives adaptation, where their gains originate, and under which shifts they remain reliable. In this paper, we take a step back from the pursuit of state-of-the-art accuracy and conduct a systematic controlled study of TTA4CLIP. We first organize existing methods into three unified paradigms according to what is updated at test time. We then introduce TTABC, an open-source TTA Benchmark for CLIP, which standardizes evaluation protocols and integrates more than 20 representative methods. Our controlled empirical analysis focuses on three key areas. First, we determine the driving factors in parameter-based methods, revealing that adaptation gains are primarily driven by test-time evidence and reliable proxies rather than heavy optimization. Second, we explore evidence utilization beyond heavy parameter tuning, showing that competitive and efficient performance can be achieved through cross- or current-sample evidence and lightweight prototype updates. Finally, we demonstrate that there is no silver bullet for TTA; no single adaptation paradigm is universally optimal, and the preferred paradigm depends on the nature of shift. We hope our benchmark and study provide a clearer understanding of the current TTA4CLIP landscape and establish a foundation for further research.

The design of attention mechanisms in long-context large language models faces an ``impossible triangle'' among KV cache size, training and prefill FLOPs, and model quality. Grouped-Query Attention (GQA) achieves strong quality but incurs high KV cache overhead; alternative approaches such as Multi-head Latent Attention (MLA) and Multi-matrix Factorization Attention (MFA) can compress the cache, but MLA suffers from quality loss while MFA substantially increases training and prefill compute. We propose Virtual Head Attention (VHA), which simultaneously improves all three vertices through two lightweight linear operations: Q Premix applies a near-identity transformation to queries within each KV group to recover virtual head diversity from halved physical query heads, and Linear Postmix fuses inter-head features via a low-rank $I + AB^\top$ residual structure. Since VHA introduces no nonlinear operations, Premix can be folded into the query projection and Postmix into the output projection at inference time, reducing the model to a standard GQA-2 architecture that directly reuses all existing GQA inference optimizations. With only two KV heads, the group size in large models readily reaches 64, fully exploiting Tensor Core parallelism during decoding. Across four scales (0.6B, 1.7B, 4B dense, and 8B-A1B MoE) with 4K-context 50B-token pretraining, VHA consistently surpasses the standard GQA baseline at every scale ($+$0.05\% to $+$0.81\%), while reducing KV cache by $4\times$ vs.\ GQA-8 at dense scales and $2\times$ vs.\ GQA-4 at the MoE scale, and lowering training and prefill FLOPs by 7.9\% (1.7B, $S{=}4096$). On an industrial 30B-A3B MoE model trained on 500B tokens at 8K context, VHA outperforms both MLA and GQA-4 by $+$0.70\,/\,$+$2.60\,\% on a 12-benchmark base-model evaluation suite, demonstrating scalability to industrial model scale. The source code will be released on GitHub.

Activation steering can change an LLM's behavior without updating its weights, but a direction intended to change safety or value stance often also shifts topic, phrasing, and task content. We study a narrower and directly testable question: can a frozen residual-stream state be equipped with an editable interface that separates semantic content from value framing well enough to reduce this collateral damage? We propose a lightweight dual-code interface with a one-way semantic$\rightarrow$value path. Semantic information may ground value recognition, while stop-gradient gating, topic de-confounding, and swap consistency discourage value supervision from rewriting the semantic factor. At inference time, we edit only the value code through a residual delta update. Across two instruction-tuned backbones, the learned interface improves the steering--damage trade-off over dense and sparse steering baselines: it preserves semantics better, reduces topic leakage and benign false refusals, and remains stable under noise, distribution shift, and recomposer controls.

Selecting the best large language model (LLM) for a fixed benchmark is often expensive, since exhaustive evaluation requires running every model on every example. Multi-armed bandit (MAB) algorithms can reduce the number of LLM calls by sequentially selecting the next model-example evaluations, thereby avoiding unnecessary computations on clearly underperforming models. Further computational savings can be achieved by predicting the evaluation scores of the various models on the set of examples. In practice, these predictions can be obtained using low-rank (LR) matrix factorization that exploits correlations in the partially observed model–example score matrix. However, such predicted evaluations are not the ground truth: they can be biased and may therefore lead to incorrect identification of the best model. In this work, we propose a principled framework that combines MAB with cheap, predicted evaluation scores without compromising statistical validity. Concretely, leveraging prediction-powered inference (PPI), we derive unbiased estimators for the performance of each model that utilize predicted scores from LR factorization to reduce variance. Importantly, this enables the construction of finite-sample valid confidence intervals in our non-i.i.d. setting, where models are selected adaptively and examples are sampled without replacement. Empirical results on real-world benchmarks demonstrate that our approach reduces the number of required evaluations, leading to meaningful savings in compute and cost, while accurately identifying the best-performing model.

The RISC-V Vector Extension (RVV) provides SIMD-style vector execution for accelerating deep learning (DL) inference on edge processors. However, existing low-rank compression methods mainly select ranks according to accuracy preservation and theoretical floating-point operation (FLOP) reduction, without considering whether the selected ranks are profitable under a target RVV effective vector length. As a result, a low-rank factorization that reduces FLOPs may still introduce tail-handling overhead or even increase the RVV instruction count under specific vector configurations. To address this problem, we propose VALOR, a low-rank restructuring framework for efficient DL deployment on RISC-V processors. Given a pretrained model and a target RVV configuration, VALOR identifies decomposable layers in the pretrained model, filters rank candidates using an RVV instruction-profitable condition, and searches for a global restructuring strategy with minimal accuracy loss. Furthermore, to avoid repeated end-to-end evaluation, VALOR uses a Hessian-aware accuracy predictor that combines layer-wise factorization error with inter-layer coupling. Experiments on Spike and a real RISC-V vector processor, SpacemiT X60, show that VALOR achieves a better accuracy-efficiency trade-off than baseline methods.

Reinforcement Learning with Verifiable Rewards (RLVR), particularly GRPO, has become the standard for eliciting LLM reasoning. However, its efficiency in exploration and difficulty adaptation remains an open challenge. In this work, we identify an implicit advantage symmetry inherent in Group Relative Advantage Estimation (GRAE) as a structural property that provides a new perspective for understanding these bottlenecks. This symmetry induces two critical limitations: (i) at the group level, strict symmetry in weights between correct and incorrect trajectories leaves unsampled action logits unchanged, thereby hindering the exploration of novel correct solution. (ii) at the sample level, the algorithm implicitly prioritizes medium-difficulty samples, remaining agnostic to the non-stationary demands of difficulty focus. Through controlled experiments, we reveal that this symmetric property is sub-optimal, yielding two pivotal insights: (i) asymmetrically down-weighting the advantages of correct trajectories encourages essential exploration but risks instability; (ii) learning efficiency can be boosted by a curriculum-like transition—prioritizing simpler samples initially before gradually shifting to complex ones. Motivated by these findings, we propose Asymmetric GRAE (A-GRAE), which dynamically modulates exploration incentives and sample-difficulty focus. Experiments across seven benchmarks demonstrate that A-GRAE consistently improves GRPO and its variants across both LLMs and MLLMs.

Spectral gradient descent (SpecGD) orthogonalizes matrix parameter updates and has inspired practical optimizers such as Muon. They often perform well in large language model training, but their dynamics remain poorly understood, especially in factorized parameterizations where the product matrix does not receive orthogonalized updates. We study such dynamics through matrix factorization (MF), where the orthogonalization is applied separately to the factor updates. We analyze spectral gradient flow (SpecGF)—a continuous-time analog of SpecGD—in the low-rank MF setting and prove "equal-rate" dynamics: all singular values grow at equal rates up to small deviations. Consequently, smaller singular values attain their target values earlier than larger ones, contrasting with the largest-first stepwise learning observed in standard gradient flow. Moreover, we prove that SpecGF in our setting converges to global minima from almost all initializations, provided the factor norms remain bounded; with $\ell_2$ regularization, we obtain global convergence. Empirically, we observe that LoRA fine-tuning with orthogonalization-based optimizers including Muon exhibit near-uniform growth in the product of LoRA adapters, consistent with the mechanism predicted by our MF analysis.

Large language models can have hidden propensities: behaviours that arise only in rare contexts. These hidden propensities pose a challenge to alignment audits, which seek to uncover problematic model behaviour but are limited in scope to scenarios anticipated by the auditor. To address this challenge, we explore _limited-parameter finetuning_ (LPF) on examples of a target behaviour. While training 3–15\% of model parameters, LPF reveals whether the model has a hidden propensity, even when standard behavioural evaluations cannot. LPF succeeds on a suite of model organisms with hidden propensities (8B-14B parameters), and also surfaces naturally occurring hidden propensities in open-weight models (27-32B parameters). LPF can thus distinguish between behaviourally equivalent models, even without knowing the context in which the propensity surfaces. LPF provides a novel affordance that may be useful for auditing frontier AI models.

Recent long-tailed recognition methods increasingly adopt two-stage fine-tuning, where a generic representation is first learned and imbalance-aware adaptation is then performed in a lightweight second stage. However, existing second-stage designs often allocate objective emphasis using delayed empirical summaries, such as historical class-wise performance or validation feedback, and commonly restrict adaptation to classifier heads. These choices may make routing unreliable when empirical feedback is noisy or skewed, and may also limit the feature plasticity required for rare categories. In this paper, we propose U-MOF, an uncertainty-guided parameter-efficient multi-objective fine-tuning framework for long-tailed recognition. U-MOF introduces lightweight residual adapters into a shared-backbone expert committee, uses entropy-based committee statistics as diagnostic routing proxies to modulate tail-oriented and smoothing-oriented objectives, and applies conflict-aware orthogonal projection to coordinate heterogeneous objective gradients. Experiments on CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 show that U-MOF improves rare-category and overall performance in controlled stage-two comparisons and remains competitive with strong published long-tailed recognition baselines, while preserving the efficiency advantages of decoupled adaptation. These results indicate that reliable stage-two long-tailed fine-tuning benefits not only from selecting appropriate objectives, but also from diagnosing when each objective should dominate and from retaining limited feature-level plasticity. An anonymized implementation is included in the Supplementary Material.

Properly tuned hyperparameters are critical for reinforcement learning algorithms to perform well, limiting their use in real world applications. AutoRL aims to to address this by nesting the reinforcement learning problem within an outer optimization loop, where a black-box optimizer tunes the reinforcement learning algorithm's hyperparameters leaving the AutoRL algorithm's hyperparameters (hyper-hyperparameters) fixed. We provide the first empirical study of an AutoRL algorithm's sensitivity with respect to its hyper-hyperparameters. Our results demonstrate that SMAC3 with Hyperband, used to tune PPO, is sensitive to its hyper-hyperparameters, albeit less so than PPO is to its hyperparameters, and this reduction comes at a significant computational cost. The per-environment tuned performance gains do not outperform random search, suggesting the reduction in sensitivity may not justify the substantial compute and engineering overhead AutoRL imposes.

Agentic workloads have emerged as a major workload for LLM inference. They differ significantly from chat-only workloads, requiring long-context processing, the ability to handle multimodal inputs, and structured multi-turn interactions with tool calling capabilities. As a result, their context exhibits structure that can carry different importance along three key axes: temporal recency to the current turn, modality such as text or image tokens, and semantic role such as user queries, tool calls, observations, or reasoning. These axes capture distinct token behaviors and lead to different sensitivities to KV-cache compression. However, existing KV-cache quantization methods are typically homogeneous or exploit only heterogeneity on a single dimension, such as temporal proximity or modality, overlooking the interactions among them. To this end, we introduce TriAxialKV, a novel mixed-precision KV-cache quantization scheme that assigns each token a triaxial tag, calibrates per-tag sensitivity, and allocates INT2/INT4 bitwidths under a fixed memory budget. We implement TriAxialKV as an end-to-end serving system, comprising calibration, mixed-precision quantization and memory management, and custom fused Triton decode kernels. When using Qwen3-VL-32B-Thinking as a computer-use agent operating the OSWorld, TriAxialKV matches the accuracy of SGLang with BF16 KV cache while supporting 4.5$\times$ KV cache size and achieving 30\% higher end-to-end throughput, when running on real GPU systems.

Many training-free sparse attention methods are effective for accelerating diffusion models. Recently, several works suggest that making sparse attention trainable can further increase sparsity while preserving generation quality. We study three key questions: (1) when do the two common masking rules, i.e., Top-k and Top-p, fail, and how can we avoid these failures? (2) why can trainable sparse attention reach higher sparsity than training-free methods? (3) what are the limitations of fine-tuning sparse attention using the diffusion loss, and how can we address them? Based on this analysis, we propose SpargeAttention2, a trainable sparse attention method that achieves high sparsity without degrading generation quality. SpargeAttention2 includes (i) a hybrid masking rule that combines Top-k and Top-p for more robust masking at high sparsity, (ii) an efficient trainable sparse attention implementation, and (iii) a distillation-inspired fine-tuning objective to better preserve generation quality during fine-tuning using sparse attention. Experiments on video diffusion models show that SpargeAttention2 reaches 95% attention sparsity and a $16.2\times$ attention speedup while maintaining generation quality, consistently outperforming prior sparse attention methods.