Deep Research systems have revolutionized how LLMs solve complex questions through iterative reasoning and evidence gathering. However, current systems remain fundamentally constrained to textual web data, overlooking the vast knowledge embedded in multimodal documents: scientific papers, technical reports, and financial documents where critical information exists in figures, tables, charts, and equations. Processing such documents demands sophisticated parsing to preserve visual semantics, intelligent chunking to maintain structural coherence, and adaptive retrieval across modalities, which are capabilities absent in existing systems. In response, we present Doc-Researcher, a unified system that bridges this gap through three integrated components: (i) deep multimodal parsing that preserves layout structure and visual semantics while creating multi-granular representations from chunk to document level, (ii) systematic retrieval architecture supporting text-only, vision-only, and hybrid paradigms with dynamic granularity selection, and (iii) iterative multi-agent workflows that decompose complex queries, progressively accumulate evidence, and synthesize comprehensive answers across documents and modalities. To enable rigorous evaluation, we introduce M4DocBench, the first benchmark for Multi-modal, Multi-hop, Multi-document, and Multi-turn deep research. Featuring 158 expert-annotated questions with complete evidence chains across 304 documents, M4DocBench tests capabilities that existing benchmarks cannot assess. Experiments demonstrate that Doc-Researcher achieves 50.6% accuracy, 3.4× better than state-of-the-art baselines, validating that effective document research requires not just better retrieval, but fundamentally deep parsing that preserve multimodal integrity and support iterative research. Our work establishes a new paradigm for conducting deep research on multimodal document collections.

Training generally capable agents in complex environments is a challenging task that involves identifying "right" environments at the training stage. Recent research has highlighted the potential of the Unsupervised Environment Design framework, which generates environment instances/levels adaptively at the frontier of the agent’s capabilities using regret measures. While regret approaches have shown great promise in generating feasible environments, they can produce difficult environments that are challenging for an RL agent to learn from. This is because regret represents the best-case (upper bound) learning potential and not the actual learning potential of an environment. To address this limitation, we propose an alternative mechanism that employs marginal benefit, focusing on the improvement (in terms of generalized performance) the agent policy gets for a given environment. The advantage of this new mechanism is that it is agent-focused (and not environment focused) and generates the "right" environments depending on the agent's policy. Additionally, to improve the generalizability of the agent, we introduce representative state diversity metric that aims to generate varied experiences for the agent. Finally, we provide detailed experimental results and ablation analysis to showcase the effectiveness of our new methods. We obtain SOTA results among RL based environment generation methods.

Wenjun Li, Dexun Li, Kuicai Dong, Cong Zhang, Hao Zhang, Weiwen Liu, Yasheng Wang, Ruiming Tang, Yong Liu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The rapid advancement of neural language models has sparked a new surge of intelligent agent research. Unlike traditional agents, large language model-based agents (LLM agents) have emerged as a promising paradigm for achieving artificial general intelligence (AGI) due to their superior reasoning and generalization capabilities. Effective planning is crucial for the success of LLM agents in real-world tasks, making it a highly pursued topic in the community. Current planning methods typically translate tasks into executable action sequences. However, determining a feasible or optimal sequence for complex tasks with multiple constraints at fine granularity, which often requires compositing long chains of heterogeneous actions, remains challenging. This paper introduces Planning with Multi-Constraints (PMC), a zero-shot methodology for collaborative LLM-based multi-agent systems that simplifies complex task planning with constraints by decomposing it into a hierarchy of subordinate tasks. Each subtask is then mapped into executable actions. PMC was assessed on two constraint-intensive benchmarks, TravelPlanner and API-Bank. Notably, PMC achieved an average 42.68% success rate on TravelPlanner, significantly higher than GPT-4 (2.92%), and outperforming GPT-4 with ReAct on API-Bank by 13.64%, showing the immense potential of integrating LLM with multi-agent systems. We also show that PMC works with small LLM as the planning core, e.g., LLaMA-3.1-8B.

Multimodal document retrieval aims to identify and retrieve various forms of multimodal content, such as figures, tables, charts, and layout information from extensive documents. Despite its increasing popularity, there is a notable lack of a comprehensive and robust benchmark to effectively evaluate the performance of systems in such tasks. To address this gap, this work introduces a new benchmark, named MMDocIR, that encompasses two distinct tasks: page-level and layout-level retrieval. The former evaluates the performance of identifying the most relevant pages within a long document, while the later assesses the ability of detecting specific layouts, providing a more fine-grained measure than whole-page analysis. A layout refers to a variety of elements, including textual paragraphs, equations, figures, tables, or charts. The MMDocIR benchmark comprises a rich dataset featuring 1,685 questions annotated by experts and 173,843 questions with bootstrapped labels, making it a valuable resource in multimodal document retrieval for both training and evaluation. Through rigorous experiments, we demonstrate that (i) visual retrievers significantly outperform their text counterparts, (ii) MMDocIR training set effectively enhances the performance of multimodal document retrieval and (iii) text retrievers leveraging VLM-text significantly outperforms retrievers relying on OCR-text.

Function calling significantly extends the application boundary of large language models (LLMs), where high-quality and diverse training data is critical for unlocking this capability. However, collecting and annotating real function-calling data is challenging, while synthetic data from existing pipelines often lack coverage and accuracy. In this paper, we present ToolACE, an automatic agentic pipeline designed to generate accurate, complex, and diverse tool-learning data, specifically tailored to the capabilities of LLMs. ToolACE leverages a novel self-evolution synthesis process to curate a comprehensive API pool of 26,507 diverse APIs. Dialogs are further generated through the interplay among multiple agents, under the guidance of a complexity evaluator. To ensure data accuracy, we implement a dual-layer verification system combining rule-based and model-based checks. We demonstrate that models trained on our synthesized data---even with only 8B parameters---achieve state-of-the-art performance, comparable to the latest GPT-4 models. Our model and a subset of the data are publicly available at https://huggingface.co/Team-ACE.

Agent decision making using Reinforcement Learning (RL) heavily relies on either a model or simulator of the environment (e.g., moving in an 8x8 maze with three rooms, playing Chess on an 8x8 board). Due to this dependence, small changes in the environment (e.g., positions of obstacles in the maze, size of the board) can severely affect the effectiveness of the policy learned by the agent. To that end, existing work has proposed training RL agents on an adaptive curriculum of environments (generated automatically) to improve performance on out-of-distribution (OOD) test scenarios. Specifically, existing research has employed the potential for the agent to learn in an environment (captured using Generalized Advantage Estimation, GAE) as the key factor to select the next environment(s) to train the agent. However, such a mechanism can select similar environments (with a high potential to learn) thereby making agent training redundant on all but one of those environments. To that end, we provide a principled approach to adaptively identify diverse environments based on a novel distance measure relevant to environment design. We empirically demonstrate the versatility and effectiveness of our method in comparison to multiple leading approaches for unsupervised environment design on three distinct benchmark problems used in literature.

Restless Multi-Armed Bandits (RMAB) is an apt model to represent decision-making problems in public health interventions (e.g., tuberculosis, maternal, and child care), anti-poaching planning, sensor monitoring, personalized recommendations and many more. Existing research in RMAB has contributed mechanisms and theoretical results to a wide variety of settings, where the focus is on maximizing expected value. In this paper, we are interested in ensuring that RMAB decision making is also fair to different arms while maximizing expected value. In the context of public health settings, this would ensure that different people and/or communities are fairly represented while making public health intervention decisions. To achieve this goal, we formally define the fairness constraints in RMAB and provide planning and learning methods to solve RMAB in a fair manner. We demonstrate key theoretical properties of fair RMAB and experimentally demonstrate that our proposed methods handle fairness constraints without sacrificing significantly on solution quality.

Influence maximization is the problem of finding a small subset of nodes in a network that can maximize the diffusion of information. Recently, it has also found application in HIV prevention, substance abuse prevention, micro-finance adoption, etc., where the goal is to identify the set of peer leaders in a real-world physical social network who can disseminate information to a large group of people. Unlike online social networks, real-world networks are not completely known, and collecting information about the network is costly as it involves surveying multiple people. In this paper, we focus on this problem of network discovery for influence maximization. The existing work in this direction proposes a reinforcement learning framework. As the environment interactions in real-world settings are costly, so it is important for the reinforcement learning algorithms to have minimum possible environment interactions, i.e, to be sample efficient. In this work, we propose CLAIM - Curriculum LeArning Policy for Influence Maximization to improve the sample efficiency of RL methods. We conduct experiments on real-world datasets and show that our approach can outperform the current best approach.