NeurIPS 2026 论文列表

Continual learning (CL) aims to adapt models to new tasks while preserving previously learned knowledge. Recently, Low-Rank Adaptation (LoRA), a representative Parameter-Efficient Fine-Tuning (PEFT) method, has gained increasing attention in CL due to its efficiency and scalability. Several LoRA-based merging methods maintain a single inference model by scaling the newly learned LoRA update with one coefficient and merging it into the previous weights after each task. However, they usually control the whole update with a shared scalar factor, although different update elements may affect previous and current tasks in different ways. To address this issue, we propose Merging Strength Modulation via Joint Loss Estimation (M²LE) It consists of Hessian Information-guided Adaptive Merging (HIM) and Task Simulation Perturbation for Merging (TSP). Instead of searching for a global merge coefficient, HIM directly solves for the target merged update under a joint objective over previous and current tasks. Solving this objective gives an analytic solution that naturally takes the form of element-wise modulation over the current task LoRA update. To improve the stability of the learned update under subsequent merging, we introduce TSP which simulates different tasks interference. Experiments on multiple benchmarks show that M²LE consistently outperforms existing PEFT-based continual learning methods. Our code is provided in the supplementary material.

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or moral dilemmas. Evaluating this fine-tuned model on six safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations.

When researchers iteratively refine ideas with large language models, do the models preserve fidelity to the original objective? We introduce DriftBench, a benchmark for evaluating constraint adherence in multi-turn LLM-assisted scientific ideation. Across 2,146 scored benchmark runs spanning seven models from five providers (including two open-weight), four interaction conditions, and 38 research briefs from 24 scientific domains, we find that iterative pressure reliably increases structural complexity and often reduces adherence to original constraints. A restatement probe reveals a dissociation between declarative recall and behavioral adherence, as models accurately restate constraints they simultaneously violate. The knows-but-violates (KBV) rate, measuring constraint non-compliance despite preserved recall, ranges from 8\% to 99\% across models. Structured checkpointing partially reduces KBV rates but does not close the dissociation, and complexity inflation persists. Human validation against blind raters confirms that the LLM judge under-detects constraint violations, making reported constraint adherence scores conservative. Sensitivity analyses confirm the findings are robust to temperature (0.7 vs.\ 1.0) and pressure type (novelty vs.\ rigor). We release all briefs, prompts, rubrics, transcripts, and scores as an open benchmark.

Machine unlearning for vision models is rapidly becoming a practical requirement, driven by privacy regulations, data errors, and the need to remove harmful or corrupted training images. Despite this, most existing approximate unlearning methods tackle the problem from a post-hoc perspective. They attempt to erase the influence of targeted samples through parameter updates that typically require access to the full training data. This creates a mismatch with real deployment scenarios where unlearning requests can be anticipated, revealing a fundamental limitation of post-hoc approaches. We motivate *unlearning by design*, a novel paradigm for approximate methods in which models are directly trained to support forgetting as an inherent architectural capability. We instantiate this idea with Machine UNlearning via KEY deletion (MUNKEY), a memory-augmented transformer that decouples instance-specific memorization from model weights. Here, unlearning corresponds to removing the instance-identifying key, enabling zero-shot forgetting without weight updates or access to the original samples or labels. Across natural image benchmarks, fine-grained visual recognition, and medical datasets, MUNKEY outperforms all post-hoc baselines. Our results establish that *unlearning by design* enables fast, deployment-oriented unlearning while preserving predictive performance.

Modelling opinion dynamics typically relies on hand-crafted local interaction rules to study emergent macroscopic phenomena such as consensus and polarisation. In contrast, multi-agent reinforcement learning (MARL) enables agents to learn such behaviours directly by optimising simple rewards. To explore the potential of MARL for opinion dynamics, we introduce a GPU-accelerated *consensus and truth-finding game* that scales to populations of up to 1000 agents, comparable to many real-world social sub-networks. To prevent unrealistic conventions, we extend other-play to general-sum social interactions. We next validate our model on a subset of the Bluesky network by recovering agent importance structures from graph topology alone via a learned attention layer, finding that *highly conforming* populations most closely match human data. In large social media networks such high levels of conforming significantly reduce collective accuracy and promote dishonest agents that lie to fit in. By contrast, small, dynamic hunter-gatherer networks are less affected; here, conformity can even improve collective agreement. This suggests a mismatch between evolved human conformity heuristics and modern social media environments as a potential contributor to misinformation.

Variational quantum algorithms offer a promising route to practical quantum advantage on near-term hardware, yet training quantum neural networks (QNNs) remains hampered by two compounding difficulties: the exponential vanishing of gradient variance known as the barren plateau, and the $\mathcal{O}(L \cdot 2^n)$ time–memory cost of differentiating through an $n$-qubit, $L$-layer circuit. We propose RLQ-Grad, a reinforcement-learning-based optimizer in which a classical policy $\pi_\phi$ (a spectrally-normalized PPO agent) learns to propose parameter updates directly, conditioned on the QNN's current parameters, loss, accuracy, and previous update. Because the surrogate gradient is emitted by a classical network rather than obtained by differentiating through the unitary $U(\theta)$, its variance is not constrained by the barren plateau concentration bound, and its per-step cost scales independent of the Hilbert-space dimension. We prove these two properties formally and verify them empirically on a hardware-efficient ansatz across four supervised benchmarks. RLQ-Grad preserves a near-flat gradient-variance curve where backpropagation, parameter-shift, and adjoint differentiation decay by 1–2 orders of magnitude, attains $128\times$–$1841\times$ wall-clock speedups and constant $\sim0.1$ MB memory at $n=12$ qubits, and improves top-1 validation accuracy by up to $+10\%$ over the strongest gradient-based baseline at every circuit scale tested.

Due to the gigapixel-scale nature of whole-slide images (WSIs), weakly supervised WSI analysis is commonly formulated as a multiple instance learning (MIL) problem, where patch-level features are aggregated into slide-level representations. However, diagnostic and prognostic evidence often arises from spatially coherent tumor microenvironment regions and their interactions, rather than isolated patches alone. Existing patch-level or static region-based methods usually overlook how tissue regions should be adaptively formed and subsequently evolved through microenvironment interactions across heterogeneous boundaries. In this paper, we propose Concept-Guided Tumor Microenvironment Evolution (TMEvolve), a reaction-diffusion-inspired framework that models WSIs as latent tumor microenvironment fields over discrete patch graphs. TMEvolve instantiates this view as a learnable graph-discretized evolution process over patch neighborhoods. It first forms adaptive soft tissue regions as coherent microenvironment units, then performs pseudo-time evolution through two complementary local dynamics: intra-region diffusion, which stabilizes latent states within coherent tissue compartments, and concept-guided boundary flux, which propagates visual feature signals and language-derived concept signals across heterogeneous region interfaces. The evolved microenvironment regions are finally aggregated for slide-level prediction. We evaluate TMEvolve on six datasets across three weakly supervised WSI tasks: survival prediction, gene expression prediction, and histological subtype classification. TMEvolve consistently improves over representative MIL methods, pathology foundation models, and concept-guided baselines. Ablation studies and visualizations further support the effectiveness and interpretability of TMEvolve, highlighting the value of dynamic region modeling and boundary interaction.

The open-source model ecosystem now contains hundreds of thousands of pretrained models, yet picking the best model for a new dataset is increasingly infeasible: new models and unbenchmarked datasets emerge continuously, leaving practitioners with no prior records on either side. Existing approaches handle only fragments of this in-the-wild setting: AutoML and transferability estimation select models from small predefined pools or require expensive per-model forward passes on the target dataset, while model routing presupposes a given candidate pool. We introduce ModelLens, a unified framework for model recommendation in the wild. Our key insight is that public leaderboard interactions, though scattered and noisy, collectively trace out an implicit atlas of model capabilities across heterogeneous evaluation settings, a signal rich enough to learn from directly. By learning a performance-aware latent space over model--dataset--metric tuples, ModelLens ranks unseen models on unseen datasets without running candidates on the target dataset. On a new benchmark of 1.62M evaluation records spanning 47K models and 9.6K datasets, ModelLens surpasses baselines that either rely on metadata alone or require running each candidate on the target dataset. Its recommended Top-K pools further improve multiple representative routing methods by up to 81\% across diverse QA benchmarks. Case studies on recently released benchmarks confirming generalization to both text and vision-language tasks.

We propose and analyze a model-based bootstrap for transition kernels in finite controlled Markov chains (CMCs) with possibly nonstationary or history-dependent control policies, a setting that arises naturally in offline reinforcement learning (RL) when the behavior policy generating the data is unknown. We establish distributional consistency of the bootstrap transition estimator in both a single long-chain regime and the episodic offline RL regime. The key technical tools are a novel bootstrap law of large numbers (LLN) for the visitation counts and a novel use of the martingale central limit theorem (CLT) for the bootstrap transition increments. We extend bootstrap distributional consistency to the downstream targets of offline policy evaluation (OPE) and optimal policy recovery (OPR) via the delta method by verifying Hadamard differentiability of the Bellman operators, yielding asymptotically valid confidence intervals for value and $Q$-functions. Experiments on the RiverSwim problem show that the proposed bootstrap confidence intervals (CIs), especially the percentile CIs, outperform the episodic bootstrap and plug-in CLT CIs, and are often close to nominal ($50\%$, $90\%$, $95\%$) coverage, while the baselines are poorly calibrated at small sample sizes and short episode lengths.

Online decision-making faces a tension between efficient inference and long-horizon reasoning. Value-based methods rely on short-horizon bootstrapping that limits the propagation of long-term information, while trajectory optimization methods plan explicitly but must solve a high-dimensional optimization problem at every decision step. We propose Generative Trajectory Planning (GTP), a model-based framework that addresses this tension by learning a generative model as a proposal distribution over action sequences. Instead of optimizing sequences from scratch, GTP samples candidate sequences from a diffusion prior and refines them using learned dynamics, reward, and value models, enabling efficient trajectory-level reasoning. Across continuous control and long-horizon manipulation tasks, GTP matches or exceeds state-of-the-art model-free, diffusion-policy, and model-based baselines, and remains stable in regimes where strong baselines degrade. Our results suggest that generative models are best understood not as policies, but as structured proposal distributions for planning, offering a new perspective on how to integrate generative modeling with model-based decision-making.

Large Language Models (LLMs) are trained to support an increasing number of languages, yet their predefined tokenizers remain a bottleneck for adapting models to lower-resource or distinct-script languages. Existing tokenizer transfer methods typically rely on semantic heuristics to initialize new embeddings, ignoring higher-layer model dynamics and limiting transfer quality. We propose Model-Aware Tokenizer Transfer (MATT), a method that incorporates model internals into the tokenizer transfer process. MATT introduces an Attention Influence Modeling (AIM) objective that distills inter-token communication patterns from a source model into a target model with a new tokenizer, providing an efficient warm-up before standard language modeling. Unlike approaches that focus solely on embedding similarity, MATT leverages attention behavior to guide embedding initialization and adaptation. Experiments across diverse linguistic settings show that MATT recovers a large fraction of the original model’s performance within a few GPU hours, outperforming heuristic baselines. These results demonstrate that incorporating model-level signals offers a practical and effective path toward robust tokenizer transfer in multilingual LLMs.

Most FDR-controlled feature selection methods are designed for coordinate-wise hypotheses, where each feature has a single weight or importance score. This abstraction fails in sequential and grouped models, where one original feature is represented by a block of sub-features, such as lags, recurrent states, or attention-based interactions. We propose a grouped-feature FDR control framework for such settings. For grouped linear models, we construct null-symmetric block-level mirror statistics with matrix-valued perturbations. For neural sequential models, we combine Permutation SHAP derivatives as model-agnostic block-level importance scores with kernel-based dependence measure. The framework is model-agnostic across network architectures, does not require specifying the covariate distribution, and reduces to Gaussian Mirror or Neural Gaussian Mirror when the block size is one. We prove FDR control for low- and high-dimensional grouped linear models and asymptotic symmetry of smoothed Permutation SHAP derivatives under fixed fitted nonlinear models. Experiments on simulated and real-world datasets show reliable FDR control and improved power under correlated grouped-feature signals.

Large language models (LLMs) increasingly act as autonomous agents that must decide when to answer directly vs. when to invoke external tools. Prior work studying adaptive tool use has largely treated tool necessity as a model-agnostic property, annotated by human or LLM judge, and mostly cover cases where the answer is obvious (e.g., fetching the weather vs.\ paraphrasing text). However, tool necessity in the wild is more nuanced due to the divergence of capability boundaries across models: a problem solvable by a strong model on its own may still require tools for a weaker one. In this work, we introduce a model-adaptive definition of tool-necessity, grounded in each model's empirical performance. Following this definition, we compare the necessity against observed tool-call behavior across four models on arithmetic and factual QA dataset, and find substantial mismatches of 26.5–54.0\% and 30.8–41.8\%, respectively. To diagnose the failure, we decompose tool use into two stages: an internal cognition stage that reflects whether a model believes a tool is necessary, and an execution stage that determines whether the model actually makes a tool-call action. By probing the LLM hidden states, we find that both signals are often linearly decodable, yet their probe directions become nearly orthogonal in the late-layer, last-token regime that drives the next-token action. By tracing the trajectory of samples in the two-stage process, we further discover that the majority of mismatch is concentrated in the cognition-to-action transition, not in cognition itself. These results reveal a knowing-doing gap in LLM tool-use: improving tool-use reliability requires not only better recognition of when tools are needed, but also better translation of that recognition into action.

A central goal of safety research is determining whether a model is misaligned. Prior work has largely focused on detecting concerning behavior, but behavior alone is not sufficient to establish misalignment: a concerning action can arise from benign causes such as confusion. This raises the problem of determining whether malign intent underlies such behavior, a process we term model incrimination. The goal of this paper is to develop effective methods for doing so. To enable this, we create a suite of six agentic environments where models exhibit concerning behavior as practice grounds, and follow a simple two-step protocol for investigating the causes behind model behavior: hypothesis generation via reading the chain of thought -- which, while not always faithful, is a rich source of hypotheses about what drives model behavior -- followed by hypothesis validation via environment interventions and additional methods as appropriate. As we do not have access to ground truth about why a model takes an action, we rely on convergent findings across independent experiments as our standard of evidence. By following our protocol, we learn effective methods for concretely determining motivations: for example, we use predictions to back out latent properties of model behavior, like the fact that Kimi K2 Thinking takes shortcuts due to a legitimate disposition towards less effortful courses of action, and that reward hacking in frontier models is strategic, while in weaker models it is not. However, some unanswered questions require further methodological development: we construct an absence-of-evidence case against Kimi K2 Thinking believing it is going against the user's wishes while taking shortcuts, but run into confounds that limit our confidence in its absence. Overall, a key takeaway is that simple methods like reading the CoT and environment interventions are highly effective. More broadly, our work shows model incrimination is a tractable empirical problem with significant room for progress, and establishes a baseline for future work.

Harmful fine-tuning can rapidly compromise safety alignment in Large Language Models, underscoring the vulnerability of deployed models to post-release modification. While prior work motivates model immunization by inducing ill-conditioned curvature to slow harmful optimization, this approach is not scalable due to the high computational cost of Hessian estimation. We refine this paradigm by extending the theoretical framework to prioritize plateau regions alongside curvature. We further introduce an immunization algorithm based on efficient Hessian estimation and bounds on the extremal singular values of the Hessian. We demonstrate the proposed immunization framework in three applications: (i) Fine-tuning-as-a-service settings where benign user data may be mixed with poisoned harmful data, (ii) open-weight settings where an adversary can fine-tune on harmful-only data with varied training dynamics, and (iii) robust unlearning, where immunized models better resist relearning attacks after unlearning. Extensive experiments confirm that our approach achieves robust defense against adversarial fine-tuning while maintaining general utility and downstream trainability.

Multimodal dataset distillation (MDD) seeks to synthesize compact surrogates from large-scale image-text corpora, reducing the high cost of pre-training. Although recent feature-matching methods have improved the efficiency of MDD, they often suffer from initialization overfitting, which distorts the cross-modal structure and severely degrades retrieval performance under unseen initializations. To mitigate this issue, we propose $\textbf{MDA-DD}$, which enhances generalization via a dynamic staggered model queue. We further identify two architectural bottlenecks in prior work: structurally redundant projections and inefficient covariance computation. For the former, we design an asymmetric dual projection to reduce redundancy while preserving balanced bidirectional cross-modal alignment. For the latter, we introduce a Second-Order Cross-Modal Moment (SOCM) objective, which efficiently captures both mean and correlation statistics across modalities. Experiments on Flickr30K and MS-COCO show that $\textbf{MDA-DD}$ consistently surpasses existing methods, yielding notable retrieval improvements (e.g., +12.9 IR@1 and +18.2 TR@1 in the 100-pair setting) and matching or exceeding baselines trained on 500 pairs using only 100 pairs.

Recursive training on synthetic data can drive generative models toward collapse, yet most explanations treat collapse as a statistical phenomenon: tail loss, variance shrinkage, or distributional drift. We develop a singular-geometric view of collapse. In a labeled Gaussian mixture abstraction of recursive self-training, we show that mode extinction induces a monotone trajectory in the real log-canonical threshold: as recursive resampling eliminates active modes, the overcomplete mixture moves through increasingly singular strata, with strictly decreasing singular complexity and growing Fisher-information nullity. We further show that this loss of identifiability begins before exact extinction: as a component's weight becomes small, Fisher curvature in its associated parameter directions shrinks proportionally, so exact singularity is the endpoint of a continuous near-nullity trajectory. Thus collapse corresponds not only to reduced distributional diversity, but also to a progressive loss of locally identifiable parameter directions. This theory separates the singular geometry of collapse from the multinomial/Wright--Fisher absorption dynamics. We also prove that, in the mixture abstraction, injecting enough real data prevents component extinction and keeps the singular-complexity trajectory from descending over a fixed horizon with high probability. We then use the theory to motivate loss-landscape diagnostics for neural recursive training, including Hessian spectra, Gauss--Newton effective rank, decoder-block curvature, feature-rank concentration, and local learning coefficient estimates. Across Gaussian mixtures, VaDE, standard VAEs, and a compact DDPM, we find that synthetic-only recursion exhibits the predicted geometric degeneration, that VaDE geometry diagnostics predict future component loss, and that real-data injection stabilizes both distributional and geometric collapse metrics.

Existing accounts of grokking explain the phenomena in terms of mechanistic frameworks such as circuit efficiency or lazy-to-rich transitions. However, despite a known dependence between grokking and model size, how model capacity shapes grokking remains an open question. We give an information-theoretic account of this relationship on the task of modular arithmetic, showing that grokking does not immediately occur when a model becomes large enough to memorise the training set, but rather emerges as the outcome of a competition between two measurable timescales: a memorisation speed $T_{\text{mem}}(P)$ and a generalisation speed $T_{\text{gen}}(P)$, both of which are functions of model parameter count $P$. Adapting the information capacity framework of Morris et al. (2025), we estimate $T_{\text{mem}}(P)$ on random-label data of equivalent complexity and $T_{\text{gen}}(P)$ on the modular task itself, and show that grokking emerges close to the parameter scale where these timescales intersect. The framework also suggests an empirical model for predicting memorisation speed given model capacity and dataset complexity, recovering the previously reported empirical observation that larger models memorise faster. Overall, we motivate the formalisation of different learning timescales as important abstractions to study when explaining how model capacity shapes grokking on algorithmic tasks.

LLM post-training is usually framed as a choice of optimizer, but KL-regularized RL also makes a hidden geometric choice: it fits the learned policy to a reward-tilted target by reverse KL. This places PPO, GRPO, RLOO, REINFORCE++, and related methods at a fixed mode-seeking point in a broader design space. We propose \textbf{Mode-Controlled Policy Optimization (MCPO)}, a geometry-aware recipe that replaces this inherited point with an $\alpha$-divergence dial over mode-seeking versus mode-covering behavior; $\alpha{=}0$ recovers the classical reverse-KL recipe. MCPO admits an off-policy objective, a practical grouped importance-weighted estimator, and a baseline-centered variant for stable training. Across mathematical reasoning, agentic transfer, search-based QA, and safety, no single geometry is uniformly best. Average-case math favors stronger concentration, while pass-based math and Search-QA favor broader support; agentic transfer moves with metric and scale; safety exposes a base-dependent frontier in which the same refusal-only signal can transfer broadly or spill into benign boundary cases depending on the reference model's mode landscape. MCPO's value is therefore not a new universal objective, but a way to expose and select the post-training geometry that RL with fixed reverse-KL geometry silently chooses.