NeurIPS 2026 论文列表

Threats and offers are fundamental to strategic interactions between agents. Competently navigating multi-agent settings requires the ability to distinguish between the two, reason about the intentions behind them, and understand how other agents assess them. With the prospect of LLM-based autonomous agents being deployed as representatives or advisors to humans in consequential domains, it is therefore important to understand how LLMs reason about threats and offers. In this paper, we propose a formal theory of threats and offers grounded in causal games, and use it to empirically investigate how current LLMs classify and reason about such proposals. We find that LLMs' threat/offer classifications align with our formal definition, with each model's threat labels tightly tracking its own judgement of whether the proposal leaves the recipient worse off. LLMs also predict their co-players' classifications accurately, though in our setting this accuracy is not clearly above what is already achieved by simply projecting their own labels onto the co-player.

Multi-task benchmarks have become a central pillar of machine learning research, yet their growing influence has incentivised benchmark gaming -- strategic actions taken to improve the leaderboard rank of a specific model. Treating datasets as voters and models as candidates, we consider benchmark-specific training -- the inclusion of benchmark data in training -- as a form of election manipulation. For any ordinal benchmark, the problem of choosing datasets to train on so that a target model becomes top-ranked corresponds to shift bribery, a class of manipulation problems from computational social choice. Leveraging this identification, we show that the benchmark-specific training problem is NP-hard under Borda count and mean win rate. Complementing this worst-case perspective, we introduce the instance-level robustness, the minimum number of datasets a model developer must include in training to top a given leaderboard, and derive expressions for it under arithmetic mean, median, mean win rate and pairwise majority. We evaluate these expressions on MMLU under HELM and on BIG-Bench Hard (BBH) under the Open LLM Leaderboard. Across both suites, mean win rate is hardest to manipulate: this gap is clear on BBH (24 tasks, 4507 models), where its median robustness is 22 tasks (92\%), compared with 13 (54\%) under arithmetic mean and 12 (50\%) under median and pairwise majority.

Low-rank adaptation (LoRA) has become the standard method for parameter-efficient fine-tuning of large pretrained models, yet theoretical explanations for why low-rank updates suffice remain incomplete. Most existing analyses rely on the Neural Tangent Kernel (NTK) regime, which linearizes the network around the pretrained weights. In this work, we take a different approach that requires no linearization: our key observation is that, once the backbone is frozen, the loss sees each adapted weight matrix only through a finite-rank linear map induced by the frozen activations. Trace-norm regularization then forces the optimizer to pick a low-rank matrix among all matrices that produce the same observed output. This yields global minimizers whose ranks are controlled by the rank of this map, which we call the \emph{bottleneck rank}, rather than the sample size. For multi-head attention, we obtain tighter rank upper bounds by exploiting two additional invariances: row-wise softmax is insensitive to additive row constants in the score matrix, and value updates are observable only after attention weighting and frozen output projection. For the factorized LoRA objective with weight decay, we prove a complementary result on the nonconvex side: after an arbitrarily small generic positive semidefinite perturbation, every first-order stationary point is rank-deficient once the LoRA rank exceeds a threshold determined by the bottleneck rank. Finally, we show that global minimizers of the perturbed objective are near-optimal for the original trace-norm problem, and we provide uniform generalization bounds governed by the spectral structure of the frozen features.

As Vision-Language Models (VLMs) are increasingly deployed as autonomous cognitive cores for embodied assistants, evaluating their privacy awareness in physical environments becomes critical. Unlike digital chatbots, these agents operate in intimate spaces, such as homes and hospitals, where they possess the physical agency to observe and manipulate privacy-sensitive information and artifacts. However, current benchmarks remain limited to unimodal, text-based representations that cannot capture the demands of real-world settings. To bridge this gap, we present ImmersedPrivacy, an interactive audio-visual evaluation framework that simulates realistic physical environments using a Unity-based simulator. ImmersedPrivacy evaluates physically grounded privacy awareness across three progressive tiers that test a model's ability to identify sensitive items in cluttered scenes, adapt to shifting social contexts, and resolve conflicts between explicit commands and inferred privacy constraints. Our evaluation of 12 state-of-the-art models reveals consistent deficits. In cluttered scenes, all models exhibit monotonic performance decay as scene complexity grows due to perceptual deficit. When social context shifts, no model exceed 65% selection accuracy. Under conflicting commands, the best model gemini-3.1-pro perfectly balances task completion and privacy preservation in only 51% of cases. These findings reveal that current VLMs in the physical world suffer from perceptual fragility and fail to let their knowledge of privacy cues govern their situated behavior. Our code and data is available at https://anonymous.4open.science/r/immersed-privacy-review-3A3C.

Post-training pruning reduces the memory and computation costs of large language models (LLMs), and N:M semi-structured sparsity provides a hardware-friendly pruning pattern for efficient inference. However, pruned models are often evaluated mainly by perplexity and average downstream accuracy, which are useful aggregate metrics but do not fully explain how pruning changes individual candidate-level decisions. We evaluate magnitude pruning, Wanda, SparseGPT, Wanda++, and our proposed MR-GOBS across multiple model families and sparsity settings. We further conduct a paired decision-level evaluation that measures true-label confidence, answer margin, prediction flips, candidate-level distribution shift, and near-boundary vulnerability. Our results show that pruning weakens decision reliability by lowering confidence in the correct answer, shrinking answer margins, and increasing correct-to-wrong flips, especially on near-boundary examples. To examine whether reconstruction-oriented improvements reliably improve decision reliability, we revisit SparseGPT through its local least-squares reconstruction formulation and refine its N:M mask-selection step. MR-GOBS uses GroupOBS, a set-level OBS reconstruction cost for jointly pruning a group of weights, to score each legal local N:M mask. It then selects masks using a minimax-regret criterion over multiple calibration views. Although MR-GOBS often improves average perplexity over SparseGPT, these gains do not always translate into better downstream accuracy or decision-level metrics. Our findings suggest that reliable evaluation of pruned LLMs requires both aggregate performance metrics and decision-level diagnostics.

Large language model (LLM)-powered assistants have recently integrated memory mechanisms that record user preferences, leading to more personalized and user-aligned responses. However, irrelevant personalized memories are often introduced into the context, interfering with the LLM's intent understanding. To comprehensively investigate the dual effects of personalization, we develop \textsc{RPEval}, a benchmark comprising a personalized intent reasoning dataset and a multi-granularity evaluation protocol. \textsc{RPEval} reveals the widespread phenomenon of irrational personalization in existing LLMs and, through error pattern analysis, illustrates its negative impact on user experience. Finally, we introduce \textsc{RP-Reasoner}, which treats memory utilization as a pragmatic reasoning process, enabling the selective integration of personalized information. Experimental results demonstrate that our method significantly outperforms carefully designed baselines on \textsc{RPEval}, and resolves 80\% of the bad cases observed in a large-scale commercial personalized assistant, highlighting the potential of pragmatic reasoning to mitigate irrational personalization. Our data is available at \url{https://anonymous.4open.science/r/RPEval-E4B0}.

While Large Language Models (LLMs) are increasingly deployed for table-related tasks, the internal mechanisms enabling them to process linearized two-dimensional structured tables remain opaque. In this work, we investigate the process of table understanding by dissecting the atomic task of cell location. Through activation patching and complementary interpretability techniques, we delineate the table understanding mechanism into a sequential three-stage pipeline: Semantic Binding, Coordinate Localization, and Information Propagation. We demonstrate that models locate the target cell via an ordinal mechanism that counts discrete delimiters to resolve coordinates, offering mechanistic evidence across diverse table formats. Furthermore, column indices are encoded within a linear subspace that allows for precise steering of model focus through vector arithmetic. Finally, we extend our analysis to the real-world HiTab dataset and show that the same three-stage mechanism persists in hierarchical real-world tables and more complex tasks.

While large language models (LLMs) appear to be increasingly capable of solving compositional tasks, it is an open question whether they do so using compositional mechanisms. In this work, we investigate how feedforward LLMs solve two-hop factual recall tasks, which can be expressed compositionally as $g(f(x))$. We first confirm that modern LLMs continue to suffer from the "compositionality gap", i.e. their ability to compute both $z = f(x)$ and $y = g(z)$ does not entail their ability to compute the composition $y = g(f(x))$. We then decode residual stream representations and identify two processing mechanisms: one which solves tasks *compositionally*, computing $f(x)$ along the way to $g(f(x))$, and one which solves them *directly*, without any detectable signature of the intermediate variable $f(x)$. Finally, we find that embedding space geometry is strongly related to which mechanism is employed, where the idiomatic mechanism is dominant when tasks are represented by translations from $x$ to $g(f(x))$ in the embedding spaces.

Tool calling, the ability to invoke external tools on demand, is central to agentic LLMs. Within an agent, the LLM decides whether and when to use these tools, greatly extending the agent’s capability boundaries. However, the mechanism that determines whether a model calls a tool or responds directly remains poorly understood. Agentic prompts are often long and heavily scaffolded, combining role instructions, tool schemas, format templates, and user requests across thousands of tokens. This creates a noisy and highly entangled context that makes it difficult to identify a single controllable variable for mechanistic analysis.To obtain such a variable, we propose a method to convert complex agentic prompts into minimal contrastive pairs, in which a single request verb determines the tool-call decision. Replacing an execution verb, such as \textit{write}, with an analysis verb, such as \textit{discuss}, reliably flips whether the model chooses to call a tool. This suggests that the choice is mediated by a compact internal state. 1,500 paired prompts across Python, Java, and C++, where 1,200 are for mechanistic analysis, and 300 are held out for evaluation, are constructed.We trace the tool-calling decision to a single vector, $\mu_\Delta$, and show that it is both causally necessary and sufficient.Transcoder analysis reveals how this vector forms: the scaffold makes tool calling the default, while analysis verbs suppress this default by activating features that signal no tool is needed. Execution verbs activate no analogous tool-use features, so $\mu_\Delta$ reflects the scaffold-induced default rather than a positive execution-verb signal. Downstream attention and MLP layers then read out this vector into the first output token. The mechanism replicates across diverse model families. Our code is available at \url{https://anonymous.4open.science/r/MI4ToolCalling}.

Despite their success in image generation, diffusion models can memorize training data, raising serious privacy and copyright concerns. Although prior work has identified empirical factors associated with memorization, such as text guidance, prompt specificity, and duplicated training examples, the mechanism by which memorization is triggered and propagated during denoising remains unclear. In this paper, we provide a theoretical explanation of how memorization occurs in text-to-image diffusion models. We show that conditional overfitting causes the conditional posterior mean to collapse to the memorized training latent, while the unconditional posterior remains close to the zero-centered data mean at the first denoising step. As a result, classifier-free guidance overestimates the clean prediction and injects an amplified memorized signal into the reverse trajectory. We further show that this signal propagates across timesteps because each guided latent is fed back into the denoising network, allowing the conditional branch to indirectly shift the unconditional branch. To characterize this process, we introduce the posterior lag, the discrepancy between conditional and unconditional posterior means. For normal prompts, this lag decreases as the two posteriors synchronize during denoising. For memorized prompts, the conditional posterior commits early to a specific training latent, while the unconditional posterior catches up only later, producing a distinctive rise-and-fall lag pattern. Our analysis provides a mechanism-level explanation of memorization and clarifies why text guidance and early denoising steps play a central role in memorized generation.

Compositional priors describe the generic properties of layered functions in deep Bayesian models, where deep neural networks with random weights are a canonical example. In the wide-network limit, the prior is a Gaussian process with a depth-dependent kernel, and its behaviour as depth grows has been extensively studied through this kernel. Here, we study another case, where each layer itself is a vector valued Gaussian process, and our aim is similarly to understand the limiting behaviour of the prior as depth grows. Previous GP work has established that for the RBF kernel and a certain range of bandwidths $r$, the prior degenerates in the limit, converging to the set of constant functions --- which is not useful as a probabilistic model. In this paper we establish several new results. First, we identify a sharp bandwidth threshold $r_c(d) = \Theta(\sqrt{d})$ above which the limit is degenerate, strengthening the earlier bounds. Second, and more importantly, we show that for $r$ below the threshold $r_c(d)$ the prior converges to a limit distribution $\pi_{\bar{Z}}$. We also prove that these distributions are non-degenerate and non-Gaussian, with non-vanishing dependence between coordinates. In contrast to the previously known degenerate regime, deep Gaussian process priors can therefore admit non-trivial limits. Empirically, we verify the threshold across a range of dimensions $d$, and demonstrate a complex multimodal behaviour of the limit distributions $\pi_{\bar{Z}}$ --- a regime that becomes increasingly narrow with $d$ and would be hard to identify without knowing the threshold.

Data augmentation is widely recognized for improving generalization in deep networks, yet its impact on the geometry of learned representations remains poorly understood. In this work, we characterize how different data augmentation strategies reshape internal representations in neural networks. Using tools from shape analysis, we embed network hidden representations into a metric space where distance is invariant to scaling, translation, rotation and reflection. We show that increasing augmentation strength leads to well-behaved trajectories in this space, and that different augmentation types steer representations in distinct directions. Moreover, we investigate how neural representation shapes are distorted along data augmentation trajectories, and show that insights from neural geometry can predict which representations provide the most improvement when ensembling models. Our results reveal shared geometric patterns across architectures and seeds, and suggest that analyzing shape-space trajectories offers a principled tool for understanding and comparing data augmentation methods.

Fluorescence spot-detection benchmarks in biology and medicine often rely on reference annotations that are incomplete, ambiguous, or inconsistently defined. In such settings, detector performance is not determined solely by model predictions; it can also depend on which faint or ambiguous signals are counted as reference positives. We formulate this issue as a reference-audit problem: instead of treating the evaluation reference as fixed ground truth, we test whether benchmark outcomes remain stable under plausible changes in reference definition and annotation completeness. Using the public FISH_spots dataset, a benchmark for fluorescence in situ hybridization (FISH) spot detection, we construct Raw, Lenient, and Strict audit references and evaluate diverse classical and learning-based detectors under a unified point-matching protocol. We further apply annotation-retention stress tests that simulate random and structured missing-label mechanisms. Across these audits, changing only the evaluation reference can alter measured F1 scores and model rankings, including the selected top-1 detector, even when aggregate rank correlation remains high. These findings suggest that model selection in fluorescence spot-detection benchmarks should be tested across plausible reference definitions, rather than based on a single reference annotation set.

Geometric modeling has become an essential paradigm in AI for Science, where data often inhabit intrinsically curved or structured spaces. However, existing manifold generative models often rely on closed-form or tractable access to global geometric primitives, such as geodesic distances, parallel transport, or spectral decompositions, which are rarely available beyond canonical geometries. To overcome this limitation, we propose Horizontal Diffusion Models (HDMs), a score-based generative framework defined on the orthonormal frame bundle of a Riemannian manifold. Instead of requiring global geometric operations, HDM lifts Euclidean diffusion processes through the Levi-Civita horizontal distribution, using only local metric and connection information while preserving intrinsic manifold geometry. This construction allows standard Euclidean score networks to be lifted into geometry-consistent and gauge-equivariant horizontal vector fields, bypassing the need for manifold-specific neural architectures. We further derive a horizontal KL objective that reduces to a Euclidean score-matching loss and analyze the curvature-dependent cost of the lift through a generalization bound. Experiments on parametric surfaces and scientific datasets demonstrate high-fidelity generation across diverse manifolds with nontrivial geometry and topology.

Learning accurate value functions plays a decisive role for reinforcement learning (RL) agents to solve long-horizon, complex tasks. Conventional temporal-difference (TD) learning objectives suffer from value-estimation bias that accumulates over the horizon, while extended-horizon modeling methods, such as $n$-step TD backups and Q-chunking, adopt a rigid, fixed-horizon value-modeling recipe that is often not flexible enough to capture complex value structures in long-horizon, multi-stage tasks. In this paper, we show that enabling value updates with dynamic horizon composition can yield a strong offline policy learning scheme. Our method, _Horizon Adaptive Offline Policy Learning via **VA**lue **ST**itching_ (**_VAST_**), replaces fixed-horizon backups with recursive, horizon-adaptive value composition. Its key ingredient is to couple value optimization with a future state- and horizon-length-conditioned **_auxiliary value function_** that is learned through direct data supervision, and a **_stitching policy_** that optimally selects the reward-maximizing horizon length and future sub-goal to achieve horizon-adaptive value stitching. This design enables direct estimation and compositional “stitching” of variable-length returns grounded in actionable sub-goal states, providing an accurate and greedily exploitable value-supervision signal for offline policy optimization. Across 50 tasks on OGBench, _VAST_ outperforms fixed-step, extended-horizon methods, and generative-value offline RL baselines, achieving strong performance particularly in high-complexity, long-horizon decision-making tasks.

Vision-language models (VLMs) show strong multimodal capabilities but still struggle with fine-grained vision-language reasoning. We find that long chain-of-thought (CoT) reasoning exposes diverse failure modes, including perception, reasoning, knowledge, and hallucination errors, which can compound across intermediate steps. However, most existing data for reinforcement learning with verifiable rewards (RLVR) does not involve complex reasoning chains grounded in visual evidence throughout, leaving these failure modes understressed during training. We therefore propose HopChain, a scalable framework that constructs an out-of-distribution (OOD) proxy task: OOD only in query format, while strengthening foundational abilities shared with downstream benchmarks to support generalizable gains. Concretely, HopChain synthesizes multi-hop vision-language reasoning chains, each a logically dependent chain of instance-grounded hops where earlier hops establish the instances, sets, or conditions needed for later hops, ending in an unambiguous numerical answer for verifiable rewards. In experiments, we train Qwen3.5-35B-A3B and Qwen3.5-397B-A17B under two RLVR settings (with and without HopChain's multi-hop data) and compare them across 24 benchmarks spanning STEM and Puzzle, General VQA, Text Recognition and Document Understanding, and Video Understanding. Although the multi-hop data is not designed for any specific benchmark, it improves 20 of 24 benchmarks on both models, indicating broad and generalizable gains. Consistently, replacing full chained queries with half-multi-hop or single-hop variants reduces the average score across five representative benchmarks from 70.4 to 66.7 and 64.3, respectively. Improvements are particularly substantial in long-CoT and ultra-long-CoT regimes, peaking at more than 50 accuracy points in the ultra-long-CoT regime. These results show that OOD proxy tasks for long-chain visual reasoning are a scalable RLVR supervision source for generalizable VLM reasoning.

Spatial transcriptomics (ST) provides spatially resolved gene expression measurements, but most 3D ST datasets are acquired as independently processed 2D sections. This sliced acquisition creates a difficult reconstruction problem: measurements are dense within each section but sparse along the axial direction, and inter-section technical variation can be confounded with true biological change. Existing coordinate-based implicit neural representations (INRs) provide a continuous modeling framework, but standard isotropic coordinate encodings do not explicitly reflect this axial-to-lateral imbalance. We present HoloGene, an anisotropic conditional INR framework for estimating continuous 3D gene-expression fields from pre-registered ST sections. HoloGene first compresses high-dimensional expression profiles into a graph-autoencoder latent manifold. It then uses pseudo-label-derived biological signature conditioning to provide coarse intra-section biological context, while depth-dependent FiLM modulation models axial variation. To reduce slice-specific covariance shifts, we introduce a signature-stratified correlation-alignment regularizer over latent features. Using held-out measured sections as cross-sectional references, HoloGene improves reconstruction accuracy and biological preservation metrics compared with general-purpose INRs and adapted spatial-transcriptomics baselines. Ablation analyses show that decoupled depth conditioning improves axial generalization, while correlation alignment provides the strongest reduction in slice-specific representation leakage. These results support anisotropic conditional modeling as a practical strategy for estimating continuous 3D expression fields from discontinuous ST sections.

Reconstructing a dynamic 4D scene along a novel camera trajectory requires both metric geometry from the source video and generative completion of disoccluded target-view regions. Feed-forward 4D reconstruction models recover increasingly comprehensive input-view geometry, including depth, point maps, and dense 3D tracks, but remain tied to the observed camera stream. Camera-controlled video diffusion models (VDMs) provide strong generative priors for novel-view synthesis, yet their camera control is typically not precise enough for metric novel-view reconstruction. We argue that comprehensive input-view 4D reconstruction is an effective geometric interface between these two capabilities: properly exposed to a VDM, it turns reconstruction outputs into camera-accurate novel-view RGB videos that remain geometrically consistent under downstream 4D reconstruction. We introduce HOLO4D, a geometry-aware video-to-video framework that controls a pretrained VDM with a hybrid 4D geometric cache built from the source video. A small set of source views covering the camera motion forms the static cache, while dynamic regions are rendered from the matching source frame. The resulting hybrid RGB-D representation is combined with explicit valid masks, target-camera ray embeddings, and dense 3D tracking tokens. Conditioned on this 4D scaffold, the diffusion backbone synthesizes the target-view RGB video. Under a common downstream 4D reconstructor, our generated videos yield target-view depth more consistent with the held-out reconstruction than that of the baselines, and an additional feature-level readout shows that intermediate VAE features already encode target-view depth. Experiments and ablations further show that progressively richer 3D/4D reconstruction signals improve camera adherence.

Scaling laws govern macroscopic resource allocation for LLMs, yet precise architectural configurations for Mixture-of-Experts (MoE) models remain guided by heuristics. Existing MoE scaling studies either incorporate MoE-specific variables into scaling formulas, causing fitted coefficients to grow rapidly without proportional experimental support, or fix all non-MoE factors, implicitly assuming global architecture does not influence local MoE scaling. We propose a reusable framework for holistic MoE architectural optimization. We first reveal that relying solely on FLOPs per token ($M$) biases MoE evaluation, as heterogeneous Attention/FFN computational densities enable parameter inflation without effective compute gains. We therefore establish a joint constraint triad of $M$, active parameters ($N_a$), and total parameters ($N$) for rigorous MoE characterization. To tame the resulting $\mathcal{O}(n^{16})$ search space, we employ mathematical decoupling: structural constraints and a rank-preserving property of the hidden dimension factorize the optimization into an $\mathcal{O}(n^3) + \mathcal{O}(n^2)$ two-phase search. Through extensive validation across 670+ MoE models spanning $10^{18}$ to $3 \times 10^{20}$ FLOPs, we derive globally applicable scaling laws mapping any compute budget to its optimal architecture. A key finding is that the near-optimal configuration band widens with scale, providing a quantitative basis for trade-offs between scaling law recommendations and engineering constraints. Our work delivers actionable blueprints for optimal MoE design under arbitrary compute budgets.