ICLR 2026 Papers

The training of large language models (LLMs) is expensive. In this paper, we study data-efficient approaches for pre-training LLMs, \ie, techniques that aim to optimize the Pareto frontier of model quality and training resource/data consumption. We seek to understand the tradeoffs associated with data selection routines based on (i) expensive-to-compute data-quality estimates, and (ii) maximization of coverage and diversity-based measures in the feature space. Our first technique, AskLLM, leverages the zero-shot reasoning capabilities of instruction-tuned LLMs to directly assess the quality of a training example. To target coverage, we propose density sampling, which models the data distribution to select a diverse sample. Testing the effect of $22$ different data curation techniques on the pre-training of T5-style of models, involving hundreds of pre-training runs and post fine-tuning evaluation tasks, we find that AskLLM and density are the best methods in their respective categories. While coverage sampling techniques often recover the performance of training on the entire dataset, training on data curated via AskLLM consistently outperforms full-data training---even when we sample only $10$\% of the original dataset, while converging up to $70$\% faster.

Universal multimodal embedding models have achieved great success in capturing semantic relevance between queries and candidates. However, current methods either condense queries and candidates into a single vector, potentially limiting the expressiveness for fine-grained information, or produce too many vectors that are prohibitively expensive for multi-vector retrieval. In this work, we introduce MetaEmbed, a new framework for multimodal retrieval that rethinks how multimodal embeddings are constructed and interacted with at scale. During training, a fixed number of learnable Meta Tokens are appended to the input sequence. At test-time, their last-layer contextualized representations serve as compact yet expressive multi-vector embeddings. Through the proposed Matryoshka Multi-Vector Retrieval training, MetaEmbed learns to organize information by granularity across multiple vectors. As a result, we enable test-time scaling in multimodal retrieval where users can balance retrieval quality against efficiency demands by selecting the number of tokens used for indexing and retrieval interactions. Extensive evaluations on the Massive Multimodal Embedding Benchmark (MMEB) and the Visual Document Retrieval Benchmark (ViDoRe) confirm that MetaEmbed achieves state-of-the-art retrieval performance while scaling robustly to models with 32B parameters. Code is available at https://github.com/facebookresearch/MetaEmbed.

We present \textsc{GenRep}, a unified image understanding and synthesis model that jointly conducts discriminative learning and generative modeling in one training session. By leveraging Monte Carlo approximation, \textsc{GenRep} distills distributional knowledge embedded in diffusion models to guide the discriminative learning for visual perception tasks. Simultaneously, a semantic-driven image generation process is established, where high-level semantics learned from perception tasks can be used to inform image synthesis, creating a positive feedback loop for mutual boosts. Moreover, to reconcile the learning process for both tasks, a gradient alignment strategy is proposed to symmetrically modify the optimization directions of perception and generation losses. These designs empower \textsc{GenRep} to be a versatile and powerful model that achieves top-leading performance on both image understanding and generation benchmarks. Code will be released after acceptance.

Large Language Models (LLMs) are increasingly being deployed as autonomous agents in real-world environments. As these deployments scale, multi-agent interactions become inevitable, making it essential to understand strategic behavior in such systems. A central open question is whether LLM agents, like reinforcement learning agents, can shape the learning dynamics and influence the behavior of others through interaction alone. In this paper, we present the first investigation of opponent shaping (OS) with LLM-based agents. Existing OS algorithms cannot be directly applied to LLMs, as they require higher-order derivatives, face scalability constraints, or depend on architectural components that are absent in transformers. To address this gap, we introduce ShapeLLM, an adaptation of model-free OS methods tailored for transformer-based agents. Using ShapeLLM, we examine whether LLM agents can influence co-players’ learning dynamics across diverse game-theoretic environments. We demonstrate that LLM agents can successfully guide opponents toward exploitable equilibria in competitive games (Iterated Prisoner’s Dilemma, Matching Pennies, and Chicken) and promote coordination and improve collective welfare in cooperative games (Iterated Stag Hunt and a cooperative version of the Prisoner’s Dilemma). Our findings show that LLM agents can both shape and be shaped through interaction, establishing opponent shaping as a key dimension of multi-agent LLM research.

Micro-expression recognition (MER) is highly challenging due to the subtle and rapid facial muscle movements and the scarcity of annotated data. Existing methods typically rely on pixel-level motion descriptors such as optical flow and frame difference, which tend to be sensitive to identity and lack generalization. In this work, we propose D-FACE, a Discrete Facial ACtion Encoding framework that leverages large-scale facial video data to pretrain an identity- and domain-invariant facial action tokenizer, for MER. For the first time, MER is shifted from relying on pixel-level motion descriptors to leveraging semantic-level facial action tokens, providing compact and generalizable representations of facial dynamics. Empirical analyses reveal that these tokens exhibit position-dependent semantics, motivating sequential modeling. Building on this insight, we employ a Transformer with sparse attention pooling to selectively capture discriminative action cues. Furthermore, to explicitly bridge action tokens with human-understandable emotions, we introduce an emotion-description-guided CLIP (EDCLIP) alignment. EDCLIP leverages textual prompts as semantic anchors for representation learning, while enforcing that the "others" category, which lacks corresponding prompts due to its ambiguity, remains distant from all anchor prompts. Extensive experiments on multiple datasets demonstrate that our method achieves not only state-of-the-art recognition accuracy but also high-quality cross-identity and even cross-domain micro-expression generation, suggesting a paradigm shift from pixel-level to generalizable semantic-level facial motion analysis.

Recent advancements in large language models (LLMs) have significantly enhanced the capabilities of collaborative multi-agent systems, enabling them to address complex challenges. However, within these multi-agent systems, the susceptibility of agents to collective cognitive biases remains an underexplored issue. A compelling example is the Mandela effect, a phenomenon where groups collectively misremember past events as a result of false details reinforced through social influence and internalized misinformation. This vulnerability limits our understanding of memory bias in multi-agent systems and raises ethical concerns about the potential spread of misinformation. In this paper, we conduct a comprehensive study on the Mandela effect in LLM-based multi-agent systems, focusing on its existence, causing factors, and mitigation strategies. We propose ManBench, a novel benchmark designed to evaluate agent behaviors across four common task types that are susceptible to the Mandela effect, using five interaction protocols that vary in agent roles and memory timescales. We evaluate agents powered by several LLMs on ManBench to quantify the Mandela effect, and analyze how different factors affect it. Moreover, we propose strategies to mitigate this effect, including prompt-level defenses (e.g., cognitive anchoring and source scrutiny) and model-level alignment-based defense, achieving an average 74.40% reduction in the Mandela effect compared to the baseline. Our findings provide valuable insights for developing more resilient and ethically aligned collaborative multi-agent systems. Code and dataset are available at https://github.com/bluedream02/Mandela-Effect.

Speech inherently contains rich acoustic information that extends far beyond the textual language. In real-world spoken communication, effective interpretation often requires integrating semantic meaning (e.g., content), paralinguistic features (e.g., emotions, speed, pitch) and phonological characteristics (e.g., prosody, intonation, rhythm), which are embedded in speech. While recent multimodal Speech Large Language Models (SpeechLLMs) have demonstrated remarkable capabilities in processing audio, their ability to perform fine-grained perception and complex reasoning in natural speech remains largely unexplored. To address this gap, we introduce MMSU, a comprehensive benchmark designed specifically for understanding and reasoning in speech. MMSU comprises 5,000 meticulously curated audio-question-answer triplets across 47 distinct tasks. Notably, linguistic theory forms the foundation of speech language understanding (SLU), yet existing benchmarks have paid insufficient attention to this fundamental aspect and fail to capture the broader linguistic picture. To ground our benchmark in linguistic principles, we systematically incorporate a wide range of linguistic phenomena, including phonetics, prosody, rhetoric, syntactics, semantics, and paralinguistics. Through a rigorous evaluation of 22 advanced SpeechLLMs, we identify substantial room for improvement in existing models. MMSU establishes a new standard for comprehensive assessment of SLLU, providing valuable insights for developing more sophisticated human-AI speech interaction systems.

Key-Value (KV) cache plays a pivotal role in accelerating inference in large language models (LLMs) by storing intermediate attention outputs, thereby avoiding redundant computation during auto-regressive generation. However, the cache's memory footprint scales linearly with sequence length, often resulting in memory bottlenecks on constrained hardware. While prior work has explored offloading KV-cache to the CPU and maintaining a reduced subset on the GPU, these approaches frequently suffer from imprecise token prioritization and degraded performance in long-generation tasks such as multi-turn dialogues and chain-of-thought reasoning. In this paper, we propose a novel KV-cache management strategy called IceCache, that integrates semantic token clustering with PagedAttention, a memory-efficient paging mechanism. By clustering semantically related tokens and organizing them into a hierarchical, dynamically updateable structure, our method improves cache hit rates and memory bandwidth utilization during CPU-GPU transfers. Experimental results show that IceCache achieves over 99\% accuracy with a 256-token budget and still maintains 97\% accuracy with only a 64-token budget, compared to the full KV-cache model. It outperforms existing baselines even while using just 25\% of the KV-cache token budget, demonstrating its superior accuracy in long-sequence scenarios.

Retrieval-Augmented Generation (RAG) has significantly enhanced LLMs by incorporating external information. However, prevailing agentic RAG approaches are constrained by a critical limitation: they treat the retrieval process as a black-box querying operation. This confines agents' actions to query issuing, hindering its ability to tackle complex information-seeking tasks. To address this, we introduce Interact-RAG, a new paradigm that elevates the LLM agent from a passive query issuer into an active manipulator of the retrieval process. We dismantle the black-box with a Corpus Interaction Engine, equipping the agent with a set of action primitives for fine-grained control over information retrieval. To further empower the agent on the entire RAG pipeline, we first develop a reasoning-enhanced workflow, which enables both zero-shot execution and the synthesis of interaction trajectories. We then leverage this synthetic data to train a fully autonomous end-to-end agent via Supervised Fine-Tuning (SFT), followed by refinement with Reinforcement Learning (RL). Extensive experiments across six benchmarks demonstrate that Interact-RAG significantly outperforms other advanced methods, validating the efficacy of our reasoning-interaction strategy.

Implicit neural representations (INRs) have recently emerged as a powerful paradigm for modeling data, offering a continuous alternative to traditional discrete signal representations. Their ability to compactly encode complex signals has led to strong performance across a wide range of computer vision tasks. In previous studies, it has been repeatedly shown that INR performance has a strong correlation with the activation functions used in its multilayer perceptrons. Although numerous competitive activation functions for INRs have been proposed, the theoretical foundations underlying their effectiveness remain poorly understood. Moreover, key challenges persist, including spectral bias (the reduced sensitivity to high-frequency signal content), limited robustness to noise, and difficulties in jointly capturing both local and global features. In this paper, we explore the underlying mechanism of INR signal representation, leveraging harmonic analysis and Chebyshev Polynomials. Through a rigorous mathematical proof, we show that modulating activation functions using a complex sinusoidal term yields better and complete spectral support throughout the INR network. To support our theoretical framework, we present empirical results over a wide range of experiments using Chebyshev analysis. We further develop a new activation function, leveraging the new theoretical findings to highlight its feasibility in INRs. We also incorporate a regularized deep prior, extracted from the signal via a task-specific model, to adjust the activation functions. This integration further improves convergence speed and stability across tasks. Through a series of experiments which include image reconstruction (with an average PSNR improvement of +5.67 dB over the nearest counterpart across a diverse image dataset), denoising (with a +0.46 dB increase in PSNR), super-resolution (with a +0.64 dB improvement over the nearest State-Of-The-Art (SOTA) method for 6X super-resolution), inpainting, and 3D shape reconstruction we demonstrate the novel proposed activation over existing state of the art activation functions.

The development of multimodal large language models (MLLMs) enables the evaluation of image quality through natural language descriptions. This advancement allows for more detailed assessments. However, these MLLM-based IQA methods primarily rely on general contextual descriptions, sometimes limiting fine-grained quality assessment. To address this limitation, we introduce a new image quality assessment (IQA) task paradigm, **grounding-IQA**. This paradigm integrates multimodal referring and grounding with IQA to realize more fine-grained quality perception, thereby extending existing IQA. Specifically, grounding-IQA comprises two subtasks: grounding-IQA-description (GIQA-DES) and visual question answering (GIQA-VQA). GIQA-DES involves detailed descriptions with precise locations (e.g., bounding boxes), while GIQA-VQA focuses on quality QA for local regions. To realize grounding-IQA, we construct a corresponding dataset, GIQA-160K, through our proposed automated annotation pipeline. Furthermore, we develop a well-designed benchmark, GIQA-Bench. The benchmark comprehensively evaluates the model grounding-IQA performance from three perspectives: description quality, VQA accuracy, and grounding precision. Experiments demonstrate that our proposed task paradigm, dataset, and benchmark facilitate the more fine-grained IQA application. Code will be made public.

In this paper, we study Riemannian zeroth-order optimization in settings where the underlying Riemannian metric $g$ is geodesically incomplete, and the goal is to approximate stationary points with respect to this incomplete metric. To address this challenge, we construct structure-preserving metrics that are geodesically complete while ensuring that every stationary point under the new metric remains stationary under the original one. Building on this foundation, we revisit the classical symmetric two-point zeroth-order estimator and analyze its mean-squared error from an intrinsic perspective, depending only on the manifold’s geometry rather than any ambient embedding. Leveraging this intrinsic analysis, we establish convergence guarantees for stochastic gradient descent (SGD) with this intrinsic estimator. Under additional suitable conditions, an $\epsilon$-stationary point under the constructed metric $g'$ also corresponds to an $\epsilon$-stationary point under the original metric $g$, thereby matching the best-known complexity in the geodesically complete setting. Empirical studies on synthetic problems confirm our theoretical findings, and experiments on a practical mesh optimization task demonstrate that our framework maintains stable convergence even in the absence of geodesic completeness.

Generative world models (WMs) can now simulate worlds with striking visual realism, which naturally raises the question of whether they can endow embodied agents with predictive perception for decision making. Progress on this question has been limited by fragmented evaluation: most existing benchmarks adopt open-loop protocols that emphasize visual quality in isolation, leaving the core issue of embodied utility unresolved, i.e., *do WMs actually help agents succeed at embodied tasks?* To address this gap, we introduce WoW!, the first open platform that benchmarks WMs in a closed-loop setting that mirrors real agent-environment interactions. WoW! provides a unified online planning strategy and a standardized action API, enabling heterogeneous WMs for decision making. We curate four closed-loop environments that rigorously evaluate diverse WMs, prioritize task success as the primary metric, and move beyond the common focus on visual quality; we also present the first data scaling law for world models in embodied settings. Our study uncovers three surprises: (1) visual quality alone does not guarantee task success—controllability matters more; (2) scaling post-training with action-observation data is more effective than upgrading the pretrained video generators; and (3) allocating more inference-time compute allows WMs to substantially improve closed-loop performance. By centering evaluation on closed-loop outcomes, WoW! establishes a new benchmark for the systematic assessment of WMs.

Large Language Models (LLMs) have been widely adopted as task planners for AI agents in sequential decision-making problems, leveraging their extensive world knowledge. However, the gap between their general knowledge and environment-specific requirements often leads to inaccurate plans. To address this, existing approaches rely on frequent LLM queries to iteratively refine plans based on immediate environmental feedback, which incurs substantial query costs. However, this refinement is typically guided by short-term environmental feedback, limiting LLMs from developing plans aligned with long-term rewards. We propose **Co**de Driven **P**lanning w**i**th Domain-Adaptive Sele**C**tor (CoPiC). Instead of relying on frequent queries, CoPiC employs LLMs to generate a diverse set of high-level planning programs, which iteratively produce and refine candidate plans. A trained domain-adaptive selector then evaluates these candidates and selects the one most aligned with long-term rewards for execution. Using high-level planning programs as planner and domain-adaptive selector as estimator, CoPiC improves planning while significantly reducing query costs. Results in ALFWorld, NetHack, and StarCraft II Unit Building show that CoPiC outperforms advanced LLM-based baselines, achieving an average (1) 19.14\% improvement in success rate and (2) 79.39\% reduction in token costs.

Large language models have achieved major advances across domains, yet training them remains extremely resource-intensive. We revisit Sign-SGD, which serves both as a memory-efficient optimizer for single-node training and as a gradient compression mechanism for distributed learning. This paper addresses a central limitation: the effective stepsize cannot be determined a priori because it relies on unknown, problem-specific quantities. We present a parameter-free Sign-SGD that removes manual stepsize selection. We analyze the deterministic single-node case, and extend the method to stochastic single-node training and multi-node settings. We also incorporate the momentum technique into our algorithms and propose a memory-efficient variant that stores only gradient signs instead of full gradients. We evaluate our methods on pre-training LLaMA models (130M and 350M) and fine-tuning a Swin Transformer (28M). Across considered tasks, the proposed methods match the performance of tuned Sign-SGD and AdamW (grid-searched stepsizes with a cosine schedule), while avoiding tuning overhead. Employing parameter-free training yields approximately $1.5\times$ end-to-end speedup compared to runs with grid-searched stepsizes.

Agentic multimodal models should not only comprehend text and images, but also actively invoke external tools, such as code execution environments and web search, and integrate these operations into reasoning. In this work, we introduce DeepEyesV2 and explore how to build an agentic multimodal model from the perspectives of data construction, training methods, and model evaluation. We observe that direct reinforcement learning alone fails to induce robust tool-use behavior. This phenomenon motivates a two-stage training pipeline: a cold-start stage to establish tool-use patterns, and reinforcement learning stage to further refine tool invocation. We curate a diverse, moderately challenging training dataset, specifically including examples where tool use is beneficial. We validate DeepEyesV2 across real-world understanding, mathematical reasoning, and search-intensive benchmarks, demonstrating that systematic tool integration enables reliable and extensible multimodal reasoning behaviour. Moreover, DeepEyesV2 exhibits task-adaptive tool invocation, tending to use image operations for perception tasks and numerical computations for reasoning tasks. Reinforcement learning further enable complex tool combinations and allowing model to selectively invoke tools based on problem context. We hope our study can provide guidance for community in developing agentic multimodal models.

Graph retrieval-augmented generation (GraphRAG) has effectively enhanced large language models in complex reasoning by organizing fragmented knowledge into explicitly structured graphs. Prior efforts have been made to improve either graph construction or graph retrieval in isolation, yielding suboptimal performance, especially when domain shifts occur. In this paper, we propose a vertically unified agentic paradigm, $\texttt{Youtu-GraphRAG}$, to jointly connect the entire framework as an intricate integration. Specifically, $(i)$ a seed graph schema is introduced to bound the automatic extraction agent with targeted entity types, relations and attribute types, also continuously expanded for scalability over unseen domains; $(ii)$ To obtain higher-level knowledge upon the schema, we develop novel dually-perceived community detection, fusing structural topology with subgraph semantics for comprehensive knowledge organization. This naturally yields a hierarchical knowledge tree that supports both top-down filtering and bottom-up reasoning with community summaries; $(iii)$ An agentic retriever is designed to interpret the same graph schema to transform complex queries into tractable and parallel sub-queries. It iteratively performs reflection for more advanced reasoning; $(iv)$ To alleviate the knowledge leaking problem in pre-trained LLM, we propose a tailored anonymous dataset and a novel 'Anonymity Reversion' task that deeply measures the real performance of the GraphRAG frameworks. Extensive experiments across six challenging benchmarks demonstrate the robustness of $\texttt{Youtu-GraphRAG}$, remarkably moving the Pareto frontier of performance and efficiency with up to 33.6% cost saving and 16.62% higher accuracy over state-of-the-art baselines. The results indicate our adaptability, allowing seamless domain transfer with minimal intervention on the schema.

Large language models (LLMs) have shown remarkable capabilities in language understanding and generation. However, such impressive capability typically comes with a substantial model size, which presents significant challenges in deployment and inference. While structured pruning of model parameters offers a promising way to reduce computational costs at deployment time, current methods primarily focus on single model pruning. In this work, we develop a novel strategy to compress models by strategically combining or merging layers from finetuned model variants, which preserves the original model's abilities by aggregating capabilities accentuated in different finetunes. We pose the optimal tailoring of these LLMs as a zero-order optimization problem, adopting a search space that supports three different operations: (1) Layer removal, (2) Layer selection from different candidate models, and (3) Layer merging. Our experiments demonstrate that this approach leads to competitive model pruning, for example, for the Llama2-13B model families, our compressed models maintain approximately 97.3\% of the original performance while removing ～25\% of parameters, significantly outperforming previous state-of-the-art methods.

We study several problems in differentially private domain discovery, where each user holds a subset of items from a shared but unknown domain, and the goal is to output an informative subset of items. For set union, we show that the simple baseline Weighted Gaussian Mechanism (WGM) has a near-optimal $\ell_1$ missing mass guarantee on Zipfian data as well as a distribution-free $\ell_\infty$ missing mass guarantee. We then apply the WGM as a domain-discovery precursor for existing known-domain algorithms for private top-$k$ and $k$-hitting set and obtain new utility guarantees for their unknown domain variants. Finally, experiments demonstrate that all of our WGM-based methods are competitive with or outperform existing baselines for all three problems.