ICLR 2026 Papers

Large Language Models (LLMs)-based text retrieval retrieves documents relevant to search queries based on vector similarities. Documents are pre-encoded offline, while queries arrive in real-time, necessitating an efficient online query encoder. Although LLMs significantly enhance retrieval capabilities, serving deeply parameterized LLMs slows down query inference throughput and increases demands for online deployment resources. In this paper, we propose LightRetriever, a novel LLM-based retriever with extremely lightweight query encoders. Our method retains a full-sized LLM for document encoding, but reduces the workload of query encoding to no more than an embedding lookup. Compared to serving a full LLM on an A800 GPU, our method achieves over a thousand times of speedup in query encoding and over 10× increase in end-to-end retrieval throughput. Extensive experiments on large-scale retrieval benchmarks show that LightRetriever generalizes well across diverse tasks, maintaining an average of 95% retrieval performance.

Modeling wind-driven object dynamics from video observations is highly challenging due to the invisibility and spatio–temporal variability of wind, as well as the complex deformations of objects. We present DiffWind, a physics-informed differentiable framework that unifies wind–object interaction modeling, video-based reconstruction, and forward simulation. Specifically, we represent wind as a grid-based physical field and objects as particle systems derived from 3D Gaussian Splatting, with their interaction modeled by the Material Point Method (MPM). To recover wind-driven object dynamics, we introduce a reconstruction framework that jointly optimizes the spatio–temporal wind force field and object motion through differentiable rendering and simulation. To ensure physical validity, we incorporate the Lattice Boltzmann Method (LBM) as a physics-informed constraint, enforcing compliance with fluid dynamics laws. Beyond reconstruction, our method naturally supports forward simulation under novel wind conditions and enable new applications such as wind retargeting. We further introduce WD-Objects, a dataset of synthetic and real-world wind-driven scenes. Extensive experiments demonstrate that our method significantly outperforms prior dynamic scene modeling approaches in both reconstruction accuracy and simulation fidelity, opening a new avenue for video-based wind–object interaction modeling. The project page is available at: [https://zju3dv.github.io/DiffWind/](https://zju3dv.github.io/DiffWind/).

Temporal grounding in long, untrimmed videos is critical for real-world video understanding, yet it remains a challenging task owing to complex temporal structures and pervasive visual redundancy. Existing methods rely heavily on supervised training with task-specific annotations, which inherently limits their scalability and adaptability due to the substantial cost of data collection and model retraining. Although a few recent works have explored training-free or zero-shot grounding, they seldom address the unique challenges posed by long videos. In this paper, we propose HiTeA (Hierarchical Temporal Alignment), a novel, training-free framework explicitly designed for long-video temporal grounding. HiTeA introduces a hierarchical temporal decomposition mechanism that structures videos into events, scenes, and actions, thereby aligning natural language queries with the most appropriate temporal granularity. Candidate segments are then matched with queries by leveraging pre-trained vision–language models (VLMs) to directly compute segment–text similarity, thereby obviating the need for any task-specific training or fine-tuning. Extensive experiments on both short- and long-video benchmarks show that HiTeA not only substantially outperforms all existing training-free methods (e.g., achieving 44.94% R\@0.1 on TACoS, representing an absolute gain of 12.4%) but also achieves competitive performance against state-of-the-art supervised baselines under stricter metrics. The code is available at https://anonymous.4open.science/r/HiTeA_code.

Agent memory shapes how Large Language Model (LLM)-powered agents, akin to the human brain, progressively refine themselves through environment interactions. Existing paradigms remain constrained: parametric memory forcibly adjusts model parameters, and retrieval-based memory externalizes experience into structured databases, yet neither captures the fluid interweaving of reasoning and memory that underlies human cognition. To address this gap, we propose MemGen, a dynamic generative memory framework that equips agents with a human-esque cognitive faculty. It consists of a \textit{memory trigger}, which monitors the agent’s reasoning state to decide explicit memory invocation, and a \textit{memory weaver}, which takes the agent's current state as stimulus to construct a latent token sequence as machine-native memory to enrich its reasoning. In this way, MemGen enables agents to recall and augment latent memory throughout reasoning, producing a tightly interwoven cycle of memory and cognition. Extensive experiments across eight benchmarks show that MemGen surpasses leading external memory systems such as ExpeL and AWM by up to $38.22\\%$, exceeds GRPO by up to $13.44\\%$, and exhibits strong cross-domain generalization ability. More importantly, we find that without explicit supervision, MemGen spontaneously evolves distinct human-like memory faculties, including planning memory, procedural memory, and working memory, suggesting an emergent trajectory toward more naturalistic forms of machine cognition.

Constrained optimization is a powerful framework for enforcing requirements on neural networks. These constrained deep learning problems are typically solved using first-order methods on their min-max Lagrangian formulation, but such approaches often suffer from oscillations and can fail to find all local solutions. While the Augmented Lagrangian method (ALM) addresses these issues, practitioners often favor dual optimistic ascent schemes (PI control) on the standard Lagrangian, which perform well empirically but lack formal guarantees. In this paper, we establish a previously unknown equivalence between these approaches: dual optimistic ascent on the Lagrangian is equivalent to gradient descent-ascent on the Augmented Lagrangian. This finding allows us to transfer the robust theoretical guarantees of the ALM to the dual optimistic setting, proving it converges linearly to all local solutions. Furthermore, the equivalence provides principled guidance for tuning the optimism hyper-parameter. Our work closes a critical gap between the empirical success of dual optimistic methods and their theoretical foundation in the single-step, first-order regime commonly used in constrained deep learning.

Accurate prediction of human behavior is essential for robust and safe human-AI collaboration. However, existing approaches for modeling people are often data-hungry and brittle because they either make unrealistic assumptions about rationality or are too computationally demanding to adapt rapidly. Our key insight is that many everyday social interactions may follow predictable patterns; efficient "scripts" that minimize cognitive load for actors and observers, e.g., "wait for the green light, then go." We propose modeling these routines as behavioral programs instantiated in computer code rather than policies conditioned on beliefs and desires. We introduce ROTE, a novel algorithm that leverages both large language models (LLMs) for synthesizing a hypothesis space of behavioral programs, and probabilistic inference for reasoning about uncertainty over that space. We test ROTE in a suite of gridworld tasks and a large-scale embodied household simulator. ROTE predicts human and AI behaviors from sparse observations, outperforming competitive baselines---including behavior cloning and LLM-based methods---by as much as 50% in terms of in-sample accuracy and out-of-sample generalization. By treating action understanding as a program synthesis problem, ROTE opens a path for AI systems to efficiently and effectively predict human behavior in the real-world.

Large language models (LLMs) are increasingly deployed in real-world applications but remain highly vulnerable to jailbreak prompts that bypass safety guardrails and elicit harmful outputs. We propose GraphShield, a graph-theoretic jailbreak detector that models information routing inside the LLM as token--layer graphs. Unlike prior defenses that rely on surface cues or costly gradient signals, GraphShield captures network-level dynamics in a lightweight and model-agnostic way by extracting multi-scale structural and semantic features that reveal jailbreak signatures. Extensive experiments on LLaMA-2-7B-Chat and Vicuna-7B-v1.5 show that GraphShield reduces attack success rates to 1.9% and 7.8%, respectively, while keeping refusal rates on benign prompts at 7.1% and 6.8%, significantly improving the robustness–utility trade-off compared to strong baselines. These results demonstrate that graph-theoretic modeling of network-level dynamics provides a principled and effective framework for robust jailbreak detection in LLMs.

Multi-output Gaussian processes (MOGPs) provide a Bayesian framework for modeling non-linear functions with multiple outputs, in which nonstationary kernels are essential for capturing input-dependent variations in observations. However, from a spectral (dual) perspective, existing nonstationary kernels inherit the inflexibility and over-parameterization of their spectral densities due to the restrictive spectral–kernel duality. To overcome this, we establish a generalized spectral–kernel duality that enables fully flexible matrix-valued spectral densities — albeit at the cost of quadratic parameter growth in the number of outputs. To achieve linear scaling while retaining sufficient expressiveness, we propose the multi-output low-rank nonstationary (MO-LRN) kernel: by modeling the spectral density through a low-rank matrix whose rows are independently parameterized by bivariate Gaussian mixtures. Experiments on synthetic and real-world datasets demonstrate that MO-LRN consistently outperforms existing MOGP kernels in regression, missing-data interpolation, and imputation tasks.

Large language models (LLMs) achieve strong downstream performance largely due to abundant supervised fine-tuning (SFT) data that imparts problem-solving capabilities. However, as applications expand, high-quality SFT data in knowledge-intensive verticals (e.g., humanities and social sciences, medicine, law, finance) is exceedingly scarce: expert curation is costly, privacy constraints are strict, and label consistency is hard to guarantee. Recent work turns to synthetic data, typically prompting a teacher model over domain documents and filtering with handcrafted rubrics. Yet, rubric design is expert-dependent and rarely transfers across domains; moreover, prevalent heuristic optimization follows a brittle loop (write rubric $\rightarrow$ synthesize $\rightarrow$ train $\rightarrow$ inspect $\rightarrow$ guess tweaks) that lacks reliable, quantitative feedback about a rubric's true contribution to downstream performance. We argue for assessing synthetic data quality through its causal impact on the target model, using this feedback to guide data generation. Inspired by classic influence functions, we repurpose an optimizer-aware estimator that uses gradient information to quantify each synthetic sample's contribution to the objective of a given target model on specific tasks. Our analysis reveals a gap: although synthetic and real samples may be close in embedding space, their influence on learning can differ substantially. Building on this insight, we propose an optimization-based synthetic data framework that adapts rubrics with target-model feedback. Instead of manually engineering domain rubrics, we supply lightweight guiding text and delegate rubric generation to a rubric-specialized model conditioned on the task; crucially, rubric (and data) selection is supervised by estimated downstream impact rather than proxy formality. Empirically, the framework yields consistent gains across domains (HSS and health), target models (e.g., Qwen and Llama families), and data generators, demonstrating broad generalization and engineering portability without task-specific tuning.

Reinforcement learning from human feedback (RLHF) with proximal policy optimization (PPO) is widely used but often yields less diverse outputs than supervised fine-tuning, suggesting an effect in which the policy’s support contracts during on-policy optimization. We formalize this “policy contraction” with the Support Retention Ratio (SRR)—the share of SFT completions that retain non-negligible probability under the RL policy—and additionally track token-entropy, Kullback–Leibler (KL) divergence to the reference, and repetition. We propose Contraction-Aware PPO (CaPPO), a minimum-norm multi-gradient update that co-optimizes reward, entropy, and KL, paired with a controller that steers exploration toward a target token entropy. On HH-RLHF, Summarize-from-Feedback, and UltraFeedback with Qwen2-7B, Qwen2.5-14B, Mistral-7B-Instruct, and Llama-3-8B-Instruct, CaPPO increases win rate by 2 to 4 points over PPO and improves diversity, gaining 0.2 to 0.3 higher SRR. The gains persist under decoding sweeps and are robust to reward scaling and critic variance. Treating reward, diversity, and stability as first-class objectives, CaPPO mitigates contraction without sacrificing alignment performance.

Prediction intervals are essential for applying machine learning models in real applications, yet most conformal prediction (CP) methods provide coverage guarantees that overlook the heterogeneity and domain knowledge that characterize modern multimodal datasets. We introduce Mixture-of-Experts Conformal Prediction (MoE-CP), a flexible and scalable framework that uses the gating probability vectors of Mixture-of-Experts (MoE) models as soft domain assignments to guide similarity-weighted conformal calibration. MoE-CP weights calibration residuals according to the similarity between gating vectors of calibration and test points, producing prediction intervals that adapt to latent subpopulations without requiring explicit domain labels. We provide theoretical justification showing that MoE-CP preserves nominal marginal validity under common similarity measures and improves conditional adaptivity when the gating captures domain structure. Empirical results on synthetic and real-world datasets demonstrate that MoE-CP yields more domain-aware, interpretable, and often tighter intervals than existing conformal baselines while maintaining target coverage. MoE-CP offers a practical route to reliable uncertainty quantification in latent heterogeneous, multi-domain environments.

In the past decade, motivated by the putative failure of naive self-play deep reinforcement learning (DRL) in adversarial imperfect-information games, researchers have developed numerous DRL algorithms based on fictitious play (FP), double oracle (DO), and counterfactual regret minimization (CFR). In light of recent results of the magnetic mirror descent algorithm, we hypothesize that simpler generic policy gradient methods like PPO are competitive with or superior to these FP-, DO-, and CFR-based DRL approaches. To facilitate the resolution of this hypothesis, we implement and release the first broadly accessible exact exploitability computations for five large games. Using these games, we conduct the largest-ever exploitability comparison of DRL algorithms for imperfect-information games. Over 7000 training runs, we find that FP-, DO-, and CFR-based approaches fail to outperform generic policy gradient methods.

The widespread adoption of AI systems in the economy hinges on their ability to generate economic value that outweighs their inference costs. Evaluating this tradeoff requires metrics that account for both performance and costs. Building on production theory, we develop an economically grounded framework for evaluating language models by combining accuracy and inference cost. We formalize cost-of-pass, the expected monetary cost of generating a correct solution. We then define the frontier cost-of-pass as the minimum cost-of-pass achievable across available models or the human-expert, using the approximate cost of hiring an expert. Our analysis reveals distinct economic insights. First, lightweight models are most cost-effective for basic quantitative tasks, large models for knowledge-intensive ones, and reasoning models for complex quantitative problems, despite higher per-token costs. Second, tracking this frontier cost-of-pass over the past year reveals significant progress, particularly for complex quantitative tasks where the cost has roughly halved every few months. Third, to trace key innovations driving this progress, we examine counterfactual frontiers—estimates of cost-efficiency without specific model classes. We find that innovations in lightweight, large, and reasoning models have been essential for pushing the frontier in basic quantitative, knowledge-intensive, and complex quantitative tasks, respectively. Finally, we assess the cost-reductions from common inference-time techniques (majority voting and self-refinement), and a budget-aware technique (TALE-EP). We find that performance-oriented methods with marginal performance gains rarely justify the costs, while TALE-EP shows some promise. Overall, our findings underscore that complementary model-level innovations are the primary drivers of cost-efficiency, and our economic framework provides a principled tool for measuring this progress and guiding deployment.

To guarantee safe and robust deployment of large language models (LLMs) at scale, it is critical to accurately assess their adversarial robustness. Existing adversarial attacks typically target harmful responses in single-point greedy generations, overlooking the inherently stochastic nature of LLMs and overestimating robustness. We show that for the goal of eliciting harmful responses, repeated sampling of model outputs during the attack complements prompt optimization and serves as a strong and efficient attack vector. By casting attacks as a resource allocation problem between optimization and sampling, we determine compute-optimal trade-offs and show that integrating sampling into existing attacks boosts success rates by up to 37\% and improves efficiency by up to two orders of magnitude. We further analyze how distributions of output harmfulness evolve during an adversarial attack, discovering that many common optimization strategies have little effect on output harmfulness. Finally, we introduce a label-free proof-of-concept objective based on entropy maximization, demonstrating how our sampling-aware perspective enables new optimization targets. Overall, our findings establish the importance of sampling in attacks to accurately assess and strengthen LLM safety at scale.

Long-context inference in large language models is bottlenecked by Key-Value (KV) cache loading during the decoding stage, where the sequential nature of generation requires repeatedly transferring the KV cache from off-chip to on-chip memory at each step. Recent architectures like Multi-Head Latent Attention (MLA) significantly reduce the KV cache size to $4.5d_h$ per token per layer while maintaining high model quality. However, when using tensor parallelism (TP) with sufficient devices for inference, MLA still decodes slower than Grouped-Query Attention (GQA) because its single latent vector cannot be sharded, forcing each device to load $4.5 d_h$ versus $2 d_h$ for GQA. In this work, we propose Multi-Head Low-Rank Attention (MLRA), a TP-friendly attention mechanism that slashes the per-device KV cache under TP to just $1.5 d_h$. Extensive experiments show that MLRA achieves state-of-the-art perplexity and downstream task performance, while also delivering a 2.8$\times$ decoding speedup over MLA.

Mixture models excel at decomposing complex, multi-modal distributions into simpler probabilistic components, but provide no insight into the conditions under which these components arise. We introduce explainable mixture models (EMM), a framework that pairs each mixture component with a human-interpretable rule over descriptive features. This enables mixtures that are not only statistically expressive but also transparently grounded in the underlying data. We formally examine the conditions under which an EMM exactly captures a target distribution and propose a scalable, differentiable learning procedure for discovering sets of rules. Experiments on synthetic and real-world datasets demonstrate that our method discovers interesting sub-populations in both univariate and multivariate settings, offering interpretable insights into the structure of complex distributions.

When model predictions inform downstream decisions, a natural question is under what conditions can the decision-makers simply respond to the predictions as if they were the true outcomes. The recently proposed notion of decision calibration addresses this by requiring predictions to be unbiased conditional on the best-response actions induced by the predictions. This relaxation of classical calibration avoids the exponential sample complexity in high-dimensional outcome spaces. However, existing guarantees are limited to linear losses. A natural strategy for nonlinear losses is to embed outcomes $y$ into an $m$-dimensional feature space $\phi(y)$ and approximate losses linearly in $\phi(y)$. Yet even simple nonlinear functions can demand exponentially large or infinite feature dimensions, raising the open question of whether decision calibration can be achieved with complexity independent of the feature dimension $m$. We begin with a negative result: even verifying decision calibration under standard deterministic best response inherently requires sample complexity polynomial in $m$. To overcome this barrier, we study a smooth variant where agents follow quantal responses. This smooth relaxation admits dimension-free algorithms: given $\mathrm{poly}(|\mathcal{A}|,1/\epsilon)$ samples and any initial predictor $p$, our introducded algorithm efficiently test and achieve decision calibration for broad function classes which can be well-approximated by bounded-norm functions in (possibly infinite-dimensional) separable RKHS, including piecewise linear and Cobb–Douglas loss functions.

Large Vision-Language Models have demonstrated exceptional performance in multimodal reasoning and complex scene understanding. However, these models still face significant hallucination issues, where outputs contradict visual facts. Recent research on hallucination mitigation has focused on retraining methods and Contrastive Decoding (CD) methods. While both methods perform well, retraining methods require substantial training resources, and CD methods introduce dual inference overhead. These factors hinder their practical applicability. To address the above issue, we propose a framework for dynamically detecting hallucination representations and performing hallucination-eliminating edits on these representations. With minimal additional computational cost, we achieve state-of-the-art performance on existing benchmarks. Extensive experiments demonstrate the effectiveness of our approach, highlighting its efficient and robust hallucination elimination capability and its powerful controllability over hallucinations. Our code will be released.

Fine-tuned Large Language Models (LLMs) encode rich task-specific features, but the form of these representations—especially within MLP layers—remains unclear. Empirical inspection of LoRA updates shows that new features concentrate in mid-layer MLPs, yet the scale of these layers obscures meaningful structure. Prior probing suggests that statistical priors may strengthen, split, or vanish across depth, motivating the need to study how neurons work together rather than in isolation. We introduce a mechanistic interpretability framework based on coalitional game theory, where neurons mimic agents in a hedonic game whose preferences capture their synergistic contributions to layer-local computations. Using top-responsive utilities and the PAC-Top-Cover algorithm, we extract stable coalitions of neurons—groups whose joint ablation has non-additive effects—and track their transitions across layers as persistence, splitting, merging, or disappearance. Applied to LLaMA, Mistral, and Pythia rerankers fine-tuned on scalar IR tasks, our method finds coalitions with consistently higher synergy than clustering baselines. By revealing how neurons cooperate to encode features, hedonic coalitions uncover higher-order structure beyond disentanglement and yield computational units that are functionally important, interpretable, and predictive across domains.