Agentic systems increasingly automate pieces of AI research. Yet turning underspecified research objectives into runnable, experimentally validated ML systems remains a central bottleneck. We study this operational setting as long-horizon ML research engineering: converting a research specification into a runnable ML system through repeated implementation, experimentation, and refinement. The central challenge is to sustain cumulative project progress across heterogeneous stages under delayed, confounded feedback. We introduce AiScientist, a multi-agent system built around thin control over thick state: a lightweight hierarchical research team coordinates through a File-as-Bus workspace that preserves decision-relevant artifacts across roles and invocations. On PaperBench, AiScientist improves over the strongest matched baselines by 9.92 and 11.15 points with Gemini-3-Flash and GLM-5, respectively. On MLE-Bench Lite, it reaches 81.82 Any Medal% under both backbones, improving over the strongest matched baselines by 4.55 and 16.67 points, and exceeding a Codex/GPT-5.5 xhigh frontier harness reference by 13.64 Any Medal points. Ablations and process analyses show that durable project state is central to later-round refinement: removing File-as-Bus lowers PaperBench score by 6.41 points and MLE-Bench Lite Any Medal% by 31.82 points. These results suggest that long-horizon AI research is not only a problem of stronger local reasoning, but a systems problem of maintaining cumulative, inspectable project progress.

Achieving mastery in real-world software engineering tasks is fundamentally bottlenecked by the scarcity of large-scale, high-quality training data. Scaling such data has been limited by the complexity of environment setup, unit-test generation, and problem statement curation. In this paper, we propose Scale-SWE, an automated, sandboxed multi-agent workflow designed to construct high-quality SWE data at scale. The system coordinates three specialized agents—for environment setup, test creation, and problem description synthesis—to process 6 million pull requests across 5.2k repositories, producing Scale-SWE-Data: 100k verified SWE instances, the largest such dataset to date. It substantially surpasses existing real-world datasets in repository diversity and reflects realistic task complexity. We further demonstrate the dataset’s utility for training by distilling 71,498 high-quality trajectories and fine-tuning Qwen-30B-A3B-Instruct to produce Scale-SWE-Agent. Our agent achieves a 64% resolve rate on SWE-Bench-Verified—a nearly three-fold improvement over the base model. Scale-SWE provides a scalable, reproducible approach for data construction to advance LLM-based software engineering.

Reinforcement learning with verifiable rewards (RLVR) enhances reasoning in large language models (LLMs) but typically depends on human-annotated answers or curated reward specifications. Label-free alternatives such as majority voting or LLM-as-a-Judge remove this cost but introduce false positives that destabilize training. We introduce JURY-RL, a label-free RLVR framework that decouples answer proposal from reward disposal: rollouts matching the plurality-voted answer receive positive reward only when that answer is verified in Lean. When verification is inconclusive, ResZero (Residual-Zero) discards the unverified plurality and redistributes a zero-mean, variance-preserving signal over the residual answers, keeping the optimization gradient stable without reinforcing unverifiable consensus. Across three backbones trained on math data, JURY-RL outperforms label-free baselines on mathematical reasoning, transfers to code generation and general benchmarks, and surpasses a majority-vote baseline in the test-time RL (TTRL) setting; it matches supervised ground-truth training on pass@1 and beats it on pass@k and response diversity.

Reinforcement learning with verifiable rewards (RLVR) has recently advanced the reasoning capabilities of large language models (LLMs). We investigate RLVR from a sample-centric perspective and introduce LPPO (Learning-Progress and Prefix-guided Optimization), a framework of progressive optimization techniques. Our work addresses a critical question: how to best leverage a small set of trusted, high-quality demonstrations, rather than simply scaling up data volume. First, motivated by how hints aid human problem-solving, we propose prefix-guided sampling, an online data augmentation method that incorporates partial solution prefixes from expert demonstrations to guide the policy, particularly for challenging instances. Second, inspired by how humans focus on important questions aligned with their current capabilities, we introduce learning-progress weighting, a dynamic strategy that adjusts each training sample’s influence based on model progression. We estimate sample-level learning progress via an exponential moving average of per-sample pass rates, promoting samples that foster learning and de-emphasizing stagnant ones. Experiments on mathematical-reasoning benchmarks demonstrate that our methods outperform strong baselines, yielding faster convergence and a higher performance ceiling, with these gains proving robust across diverse model architectures, scales, and reinforcement learning optimizers.

Retrieval-Augmented Generation (RAG) systems have emerged as a pivotal methodology for enhancing Large Language Models (LLMs) through the dynamic integration of external knowledge. To further improve RAG’s flexibility, Agentic RAG introduces autonomous agents into the workflow. However, Agentic RAG faces several challenges:(1) the success of each step depends on both high-quality planning and accurate search,(2) the lack of supervision for intermediate reasoning steps, and(3) the exponentially large candidate space for planning and searching.To address these challenges, we propose DecoupleSearch, a novel framework that decouples planning and search processes using dual value models, enabling independent optimization of plan reasoning and search grounding. Our approach constructs a reasoning tree, where each node represents planning and search steps. We leverage Monte Carlo Tree Search to assess the quality of each step. During inference, Hierarchical Beam Search iteratively refines planning and search candidates with dual value models. Extensive experiments across policy models of varying parameter sizes, demonstrate the effectiveness of our method.

Despite the remarkable capabilities of large language models (LLMs) in various reasoning tasks, they still struggle with table reasoning tasks, particularly in maintaining consistency throughout multi-step reasoning processes. While existing approaches have explored various decomposition strategies, they often lack effective mechanisms to identify and correct errors in intermediate reasoning steps, leading to cascading error propagation. To address these issues, we propose Table-Critic, a novel multi-agent framework that facilitates collaborative criticism and iterative refinement of the reasoning process until convergence to correct solutions. Our framework consists of four specialized agents: a Judge for error identification, a Critic for comprehensive critiques, a Refiner for process improvement, and a Curator for pattern distillation. To effectively deal with diverse and unpredictable error types, we introduce a self-evolving template tree that systematically accumulates critique knowledge through experience-driven learning and guides future reflections. Extensive experiments have demonstrated that Table-Critic achieves substantial improvements over existing methods, achieving superior accuracy and error correction rates while maintaining computational efficiency and lower solution degradation rate.

Direct Preference Optimization (DPO) using an implicit reward model has proven to be an effective alternative to reinforcement learning from human feedback (RLHF) for fine-tuning preference aligned large language models (LLMs). However, the overall preference annotations of responses do not fully capture the fine-grained quality of model outputs in complex multi-step reasoning tasks, such as mathematical reasoning. To address this limitation, we introduce a novel algorithm called Step-level Value Preference Optimization (SVPO). Our approach employs Monte Carlo Tree Search (MCTS) to automatically annotate step-level preferences for multi-step reasoning. Furthermore, from the perspective of learning-to-rank, we train an explicit value model to replicate the behavior of the implicit reward model, complementing standard preference optimization. This value model enables the LLM to generate higher reward responses with minimal cost during inference. Experimental results demonstrate that our method achieves state-of-the-art performance on both in-domain and out-of-domain mathematical reasoning benchmarks.

Elucidating the reasoning process with structured explanations from question to answer is crucial, as it significantly enhances the interpretability, traceability, and trustworthiness of question-answering (QA) systems. However, structured explanations demand models to perform intricately structured reasoning, which poses great challenges. Most existing methods focus on single-step reasoning through supervised learning, ignoring logical dependencies between steps. Moreover, existing reinforcement learning (RL) based methods overlook the structured relationships, underutilizing the potential of RL in structured reasoning. In this paper, we propose SEER, a novel method that maximizes a structure-based return to facilitate structured reasoning and explanation. Our proposed structure-based return precisely describes the hierarchical and branching structure inherent in structured reasoning, effectively capturing the intricate relationships between different reasoning steps. In addition, we introduce a fine-grained reward function to meticulously delineate diverse reasoning steps. Extensive experiments show that SEER significantly outperforms state-of-the-art methods, achieving an absolute improvement of 6.9% over RL-based methods on EntailmentBank, a 4.4% average improvement on STREET benchmark, and exhibiting outstanding efficiency and cross-dataset generalization performance.

The large language models have achieved superior performance on various natural language tasks. One major drawback of such approaches is they are resource-intensive in fine-tuning new datasets. Soft-prompt tuning presents a resource-efficient solution to fine-tune the pre-trained language models (PLMs) while keeping their weight frozen. Existing soft prompt methods mainly focus on designing the input-independent prompts that steer the model to fit the domain of the new dataset. Those methods often ignore the fine-grained information about the task and context of the text. In this paper, we propose a multi-level prompt tuning (MPrompt) method for machine reading comprehension. It utilizes prompts at task-specific, domain-specific, and context-specific levels to enhance the comprehension of input semantics at different granularities. We also propose an independence constraint to steer each domain-specific prompt to focus on information within its domain to avoid redundancy. Moreover, we present a prompt generator that incorporates context-related knowledge in the prompt generation to enhance contextual relevancy. We conducted extensive experiments on 12 benchmarks of various QA formats and achieved an average improvement of 1.94% over the state-of-the-art methods.