Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists’ First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Static benchmarks often conflate memorization with reasoning, failing to capture the dynamic nature of world knowledge. We present LiveSearchBench, a scalable pipeline that automatically constructs snapshot-grounded benchmarks from knowledge graph differentials, in which every answer is unique and SPARQL-verifiable against a target snapshot. Unlike prior template-based approaches, our method leverages topology-aware synthesis to generate multi-constraint compositional questions with unique answers guaranteed via SPARQL validation. Evaluating 1,000 questions across three difficulty levels on both open-weight and frontier models, we expose a pronounced Recency Gap on facts that post-date a model’s training cutoff, particularly for compositional queries; retrieval dependence is thus an empirical property we measure, not a formal guarantee we claim. Through Oracle and Wikidata-as-corpus analyses, we disentangle dual failure modes: retrieval coverage limitations for single-hop questions and reasoning failures for compositional queries even with ground-truth evidence. LiveSearchBench provides both a rigorous construction methodology and a continuously updatable testbed for evaluating evidence integration under evolving knowledge.

We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bias, where agents bypass visual tools in favor of textual search, and (2) parametric knowledge leakage, where models rely on internal memory rather than genuine tool-augmented execution. To address these challenges, we propose Video-DeepResearch, featuring a decoupled perception-exploration pipeline with stage-wise tool unlocking that compels exhaustive cross-frame visual grounding prior to web retrieval. Our framework adopts a two-stage training recipe—supervised fine-tuning followed by Group Relative Policy Optimization (GRPO)—enabling autonomous exploration that breaks the imitation-learning ceiling. Furthermore, we curate VideoDR-Bench, a human-AI collaborative benchmark comprising 200 complex, multi-hop VQA instances. Empirical results demonstrate that our Video-DeepResearch-35B-A3B establishes a new state-of-the-art of 64.0% average accuracy, surpassing proprietary Claude-4.5-Sonnet (59.0%) by 5.0 points and significantly outperforming GPT-5 (52.5%) and Gemini 2.5 Pro (57.5%). The 30B-A3B variant achieves 59.3%, competitive with Claude-4.5-Sonnet and demonstrating the effectiveness of our training paradigm even at compact scale.

Large vision–language model (LVLM)-based web agents are emerging as powerful automation tools but face severe security risks in real-world deployment. Existing benchmarks offer limited coverage, typically isolating user-level prompts from environmental threats, thus failing to capture the full spectrum of vulnerabilities. To address this, we present SecureWebArena, the first holistic security benchmark for web agents. SecureWebArena features a unified suite of six realistic web environments with 2,970 adversarial trajectories, covering a structured taxonomy of six attack vectors that span both user-level and environment-level manipulations. Crucially, we introduce a multi-layered evaluation protocol that dissects agent failures across internal reasoning, behavioral execution, and task outcomes, enabling fine-grained risk analysis beyond simple success metrics. Experiments on 9 representative LVLMs reveal universal vulnerabilities to subtle manipulations and uncover significant trade-offs between model specialization and security. SecureWebArena establishes a rigorous foundation for advancing the development of trustworthy web agents.

Recent works have increasingly applied Large Language Models (LLMs) as agents in financial stock market simulations to test if micro-level behaviors aggregate into macro-level phenomena. However, a crucial question arises: Do LLM agents’ behaviors align with real market participants? This alignment is key to the validity of simulation results. To explore this, we select a financial stock market scenario to test behavioral consistency. Investors are typically classified as fundamental or technical traders, but most simulations fix strategies at initialization, failing to reflect real-world trading dynamics. In this work, we assess whether agents’ strategy switching aligns with financial theory, providing a framework for this evaluation. We operationalize four behavioral-finance drivers—loss aversion, herding, wealth differentiation, and price misalignment—as personality traits set via prompting and stored long-term. In year-long simulations, agents process daily price-volume data, trade under a designated style, and reassess their strategy every 10 trading days. We introduce four alignment metrics and use Mann–Whitney U tests to compare agents’ style-switching behavior with financial theory. Our results show that recent LLMs’ switching behavior is only partially consistent with behavioral-finance theories, highlighting the need for further refinement in aligning agent behavior with financial theory.

We propose Process-Aware Policy Optimization (PAPO), a method that integrates process-level evaluation into Group Relative Policy Optimization (GRPO) through **decoupled advantage normalization**, to address two limitations of existing reward designs. Outcome reward models (ORM) evaluate only final-answer correctness, treating all correct responses identically, and gradually lose the advantage signal as groups become uniformly correct. Rubric-based process reward models (PRM) offer richer supervision, differentiating quality among correct responses and restoring advantage diversity, but using their scores as rewards triggers verbosity-driven reward hacking and accuracy collapse. PAPO resolves both by composing the advantage from an outcome advantage A_out, derived from outcome reward r^out over all responses, and a process advantage A_proc, derived from process reward r^proc exclusively over the correct subset. A_out anchors correctness, and correct subset normalization on A_proc rules out verbosity-driven hacking while sustaining the gradient on all-correct groups where ORM exhausts. PAPO consistently outperforms ORM across five model configurations and six benchmarks (reaching 51.3% vs. 46.3% on OlympiadBench for Qwen2.5-7B while continuing to improve as ORM plateaus and declines), with analogous gains when composed with mainstream GRPO variants.

While large language models (LLMs) perform strongly after adaptation within individual scientific domains, their learning dynamics in multi-disciplinary contexts remain poorly understood, despite the promise of improved generalization and broader applicability through cross-domain knowledge synergy. We present a systematic study of multi-discipline LLM fine-tuning across five scientific disciplines and compare the learning dynamics of full fine-tuning, LoRA, LoRA-MoE, and LoRA composition. In our naturally imbalanced setting, we show that multi-disciplinary learning is substantially more variable than single-discipline training and distill four empirical patterns: (1) Balance-then-Diversity: increasing scarce-domain exposure helps, but broad source-pool diversity must be preserved; (2) Merge-then-Align: general instruction data restores instruction-following behavior needed for effective joint adaptation; (3) Optimize-then-Scale: increasing trainable capacity offers limited gains without prior design optimization; and (4) Share-then-Specialize: asymmetric LoRA-MoE improves multi-discipline performance through a shared low-rank projection. Together, these patterns form a practical recipe for principled multi-discipline fine-tuning and provide actionable guidance for developing generalizable scientific LLMs.

Multi-agent LLM systems often fail not for lack of candidate answers, but because they have no persistent mechanism for deciding which incompatible claim should currently be trusted. Majority vote, debate, and judge-based selection choose an output without recording which claim wins, which is contested, or why a later update supersedes it. We present LatticeMind, a conflict-aware structured memory that handles contradiction at write time. It maintains explicit item status, applies cheap symbolic conflict checks, and routes the flagged cases to an LLM reconciler. On a label-blind ConflictBank evaluation that removes source-name hints, over a uniform draw of 200 examples from the full benchmark pool, LatticeMind reaches 0.965 accuracy versus 0.525 for the strongest baseline, with every gap significant at p<10^-24 by paired McNemar test. Ablations show that removing the checker or the reconciler costs 12 to 14 points. Our claims are scoped to credibility conflicts, where one claim should defeat another: the coordination branch of the design is a safety override that we do not validate empirically, and all prompt-based comparisons are conditioned on a single backbone. On three secondary planning benchmarks and one synthetic scheduling diagnostic the picture is mixed: LatticeMind beats naive merge on three of four, but does not replace deliberation methods on tasks rewarding iterative search.

Test-time Scaling (TTS) has been demonstrated to significantly enhance the reasoning capabilities of Large Language Models (LLMs) during the inference phase without altering model parameters. However, existing TTS methods are largely independent, implying that LLMs have not yet evolved to progressively learn how to scale more effectively. With the objective of evolving LLMs to learn “how to scale test-time computation,” we propose LatentEvolve, a self-evolving latent TTS framework inspired by the complementary learning system (CLS) theory. Analogous to the human brain’s dual system of a fast-recall hippocampus and a slow-consolidating neocortex, LatentEvolve comprises two evolutionary components: daytime scaling, which rapidly retrieves historical latent representations to better guide current LLM reasoning; and nighttime scaling, which integrates past latent optimizations in a manner akin to the human brain’s consolidation of experiences during sleep. The alternation of daytime and nighttime processes facilitates a fast and slow evolution of LLM TTS, mirroring human cognitive dynamics in a fully unsupervised manner. Extensive experiments across eight benchmarks and five model backbones demonstrate that our LatentEvolve surpasses state-of-the-art TTS methods such as LatentSeek and TTRL by up to 13.33% and exhibits exceptional cross-domain and cross-backbone generalization.

Agentic learning increasingly hinges on interaction, yet real-world experience is expensive, limited, and often irreversible at inference time. World models promise to mitigate these limitations, but it remains unclear whether large language models can actually serve as reliable world models, and deliver concrete benefits to downstream agents. We investigate these questions in text-based environments, a controlled testbed that reframes language modeling as next-state prediction under interaction. We propose a three-level framework to evaluate LLM-based world models: (i) fidelity and consistency, (ii) scalability and robustness, and (iii) agent utility. Across five representative environments, we show that sufficiently trained world models capture coherent environment dynamics, scale predictably with data and model capacity, and unlock tangible agent improvements—for example, action verification boosts GPT-4o by 5.5% on WebShop, and warm-started RL achieves a 15% gain on SciWorld. Crucially, these benefits hinge on behavioral coverage and environment complexity, sharply characterizing when world modeling meaningfully advances agent learning.

While scaling laws for large language models (LLMs) during pre-training have been extensively studied, their behavior under reinforcement learning (RL) post-training remains largely unexplored. This paper investigates the scaling behavior of Large Language Model (LLM) reinforcement learning post-training, focusing on mathematical reasoning. Through experiments across the Qwen2.5 series (0.5B to 72B), we characterize how model scale, data, and compute interact. Our analysis yields four key findings: 1. Larger models consistently demonstrate superior compute and data efficiency. 2. The relationship between model performance and training resources follows a predictive power-law across both base and instruction-tuned models. 3. RL learning efficiency exhibits a latent saturation trend with increasing model scale. 4. In data-constrained regimes, performance is primarily driven by the total volume of training data rather than sample uniqueness. These results offer practical guidelines for scaling reasoning capabilities through reinforcement learning post-training.

Tool-using agents based on Large Language Models (LLMs) excel in tasks such as mathematical reasoning and multi-hop question answering. However, in long trajectories, agents often trigger excessive and low-quality tool calls, increasing latency and degrading inference performance, making managing tool-use behavior challenging. In this work, we conduct entropy-based pilot experiments and observe a strong positive correlation between entropy reduction and high-quality tool calls. Building on this finding, we propose using entropy reduction as a supervisory signal and design two reward strategies to address the differing needs of optimizing tool-use behavior. Sparse outcome rewards provide coarse, trajectory-level guidance to improve efficiency, while dense process rewards offer fine-grained supervision to enhance performance. Experiments across diverse domains show that both reward designs improve tool-use behavior: the former reduces tool calls by 72.07% compared to the average of baselines, while the latter improves performance by 22.27%. These results position entropy reduction as a key mechanism for enhancing tool-use behavior, enabling agents to be more adaptive in real-world applications.

Multi-agent systems have emerged as a promising approach for enhancing the reasoning capabilities of large language models in complex problem-solving. However, current MAS frameworks are limited by poor flexibility and scalability, with underdeveloped optimization strategies. To address these challenges, we propose ReSo, which integrates task graph generation with a reward-driven two-stage agent selection process. The core of ReSo is the proposed Collaborative Reward Model, which can provide fine-grained reward signals for MAS cooperation for optimization. We also introduce an automated data synthesis framework for generating MAS benchmarks, without human annotations. Experimentally, ReSo matches or outperforms existing methods. ReSo achieves 33.7% and 32.3% accuracy on Math-MAS and SciBench-MAS SciBench, while other methods completely fail. The code and data are available at Reso.

The rapid advancement of scientific progress requires innovative tools that can accelerate knowledge discovery. Although recent AI methods, particularly large language models (LLMs), have shown promise in tasks such as hypothesis generation and experimental design, they fall short of replicating the collaborative nature of real-world scientific practices, where diverse experts work together in teams to tackle complex problems. To address the limitations, we propose an LLM-based multi-agent system, i.e., Virtual Scientists (VIRSCI), designed to mimic the teamwork inherent in scientific research. VIRSCI organizes a team of agents to collaboratively generate, evaluate, and refine research ideas. Through comprehensive experiments, we demonstrate that this multi-agent approach outperforms the state-of-the-art method in producing novel scientific ideas. We further investigate the collaboration mechanisms that contribute to its tendency to produce ideas with higher novelty, offering valuable insights to guide future research and illuminating pathways toward building a robust system for autonomous scientific discovery. The code is available at https://github.com/open-sciencelab/Virtual-Scientists.

Ensuring the trustworthiness of large language models (LLMs) is crucial. Most studies concentrate on fully pre-trained LLMs to better understand and improve LLMs’ trustworthiness. In this paper, to reveal the untapped potential of pre-training, we pioneer the exploration of LLMs’ trustworthiness during this period, focusing on five key dimensions: reliability, privacy, toxicity, fairness, and robustness. To begin with, we apply linear probing to LLMs. The high probing accuracy suggests that LLMs in early pre-training can already distinguish concepts in each trustworthiness dimension. Therefore, to further uncover the hidden possibilities of pre-training, we extract steering vectors from a LLM’s pre-training checkpoints to enhance the LLM’s trustworthiness. Finally, inspired by the theoretical result that mutual information estimation is bounded by linear probing accuracy, we also probe LLMs with mutual information to investigate the dynamics of trustworthiness during pre-training. We are the first to observe a similar two-phase phenomenon: fitting and compression. This research provides an initial exploration of trustworthiness modeling during LLM pre-training, seeking to unveil new insights and spur further developments in the field.