arXiv subjects

Jianbo Lin

Publications and source records attributed to Jianbo Lin.

7 recordsLinked to original sources

arXiv2026

From Brewing to Resolution: Tracing the Internal Lifecycle of Code Reasoning in LLMs

Standard accuracy metrics cannot explain why LLMs handle variable tracking but fail on semantically equivalent loops. We study an internal lifecycle of code reasoning in which models first brew the answer, making it linearly recoverable many layers before it becomes self-decodable, and then diverge into one of four resolution outcomes: Resolved, Overprocessed, Misresolved, or Unresolved. Understanding this lifecycle matters because similar task accuracies can mask fundamentally different failure modes that surface-level evaluation cannot detect. We introduce a dual diagnostic framework pairing layer-wise linear probing with Context-Stripped Decoding (CSD) and apply it to six code-reasoning task families across 16 models spanning Qwen, Llama, and DeepSeek architectures. All four outcomes carry substantial mass in every task family: overall Resolved is only 41.5%, with multiple tasks below 30%. Controlled sweeps over structure, depth, and operators expose task-specific failure bottlenecks: Function Call Resolved plunges from 61.1% to 2.5% as call depth increases from one to three. Across architectures and scales, the brewing scaffold remains stable, with normalized brewing duration 24-42% across all 16 models, while resolution success varies with capability. This indicates that the scaffold is a stable empirical regularity across the tested decoder-only Transformer families, whereas resolution success covaries with capability, scale, and training. Code: https://github.com/euyis1019/llm-brewing

cs.AI↗

arXiv2026

ICRL: Learning to Internalize Self-Critique with Reinforcement Learning

Large language model-based agents make mistakes, yet critique can often guide the same model toward correct behavior. However, when critique is removed, the model may fail again on the same query, indicating that it has not internalized the critique's guidance into its underlying capability. Meanwhile, a frozen critic cannot improve its feedback quality over time, limiting the potential for iterative self-improvement. To address this, we propose learning to internalize self-critique with reinforcement learning(ICRL), a novel framework that jointly trains a solver and a critic from a shared backbone to convert critique-induced success into unassisted solver ability. The critic is rewarded based on the solver's subsequent performance gain, incentivizing actionable feedback. To address the distribution shift between critique-conditioned and critique-free behavior, ICRL introduces a distribution-calibration re-weighting ratio that selectively transfers critique-guided improvements compatible with the solver's own prompt distribution. Additionally, a role-wise group advantage estimation stabilizes joint optimization across the two roles. Together, these mechanisms ensure that the solver learns to improve itself without external critique, rather than becoming dependent on critique-conditioned behavior. We evaluate ICRL on diverse benchmarks spanning agentic and mathematical reasoning tasks, using Qwen3-4B and Qwen3-8B as backbones. Results show consistent improvements, with average gains of 6.4 points over GRPO on agentic tasks, and 7.0 points on mathematical reasoning. Notably, the learned 8B critic is comparable to 32B critics while using substantially fewer tokens. The code is available at https://github.com/brick-pid/ICRL.

cs.AI↗

arXiv2025

Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL

Recent advances in large language models (LLMs) and multi-agent systems have demonstrated remarkable capabilities in complex problem-solving tasks such as deep research, vibe coding, and mathematical reasoning. However, most existing multi-agent systems are built upon manual prompt/workflow engineering with sophisticated agent frameworks, making them computationally inefficient, less capable, and can not benefit from data-centric learning. In this work, we introduce Chain-of-Agents (CoA), a novel paradigm of LLM reasoning that enables native end-to-end complex problem-solving in the same way as a multi-agent system (i.e., multi-turn problem solving with multiple tools and multiple agents) within one model. In chain-of-agents problem-solving, the model dynamically activates different tool agents and role-playing agents to simulate multi-agent collaboration in an end-to-end fashion. To elicit end-to-end chain-of-agents problem-solving abilities in LLMs, we introduce a multi-agent distillation framework to distill state-of-the-art multi-agent systems into chain-of-agents trajectories for agentic supervised fine-tuning. We then use agentic reinforcement learning on verifiable agentic tasks to further improve the models' capabilities on chain-of-agents problem solving. We call the resulting models Agent Foundation Models (AFMs). Our empirical studies demonstrate that AFM establishes new state-of-the-art performance across diverse benchmarks in both web agent and code agent settings. We make the entire research, including the model weights, code for training and evaluation, and the training data, fully open-sourced, which offers a solid starting point for future research on agent models and agentic RL.

cs.AI↗

arXiv2021

Unsupervised learning-based structural analysis: Search for a characteristic low-dimensional space by local structures in atomistic simulations

Owing to the advances in computational techniques and the increase in computational power, atomistic simulations of materials can simulate large systems with higher accuracy. Complex phenomena can be observed in such state-of-the-art atomistic simulations. However, it has become increasingly difficult to understand what is actually happening and mechanisms, for example, in molecular dynamics (MD) simulations. We propose an unsupervised machine learning method to analyze the local structure around a target atom. The proposed method, which uses the two-step locality preserving projections (TS-LPP), can find a low-dimensional space wherein the distributions of datapoints for each atom or groups of atoms can be properly captured. We demonstrate that the method is effective for analyzing the MD simulations of crystalline, liquid, and amorphous states and the melt-quench process from the perspective of local structures. The proposed method is demonstrated on a silicon single-component system, a silicon-germanium binary system, and a copper single-component system.

cond-mat.mtrl-sci↗

arXiv2020

Large scale and linear scaling DFT with the CONQUEST code

We survey the underlying theory behind the large-scale and linear scaling DFT code, Conquest, which shows excellent parallel scaling and can be applied to thousands of atoms with exact solutions, and millions of atoms with linear scaling. We give details of the representation of the density matrix and the approach to finding the electronic ground state, and discuss the implementation of molecular dynamics with linear scaling. We give an overview of the performance of the code, focussing in particular on the parallel scaling, and provide examples of recent developments and applications.

physics.comp-ph↗

arXiv2019

Highly accurate local basis sets for large-scale DFT calculations in CONQUEST

Given the widespread use of density functional theory (DFT), there is an increasing need for the ability to model large systems (beyond 1,000 atoms). We present a brief overview of the large-scale DFT code Conquest, which is capable of modelling such large systems, and discuss approaches to the generation of consistent, well-converged pseudo-atomic basis sets which will allow such large scale calculations. We present tests of these basis sets for a variety of materials, comparing to fully converged plane wave results using the same pseudopotentials and grids.

cond-mat.mtrl-sci↗

arXiv2018

Machine learning forces trained by Gaussian process in liquid states: Transferability to temperature and pressure

We study a generalization performance of the machine learning (ML) model to predict the atomic forces within the density functional theory (DFT). The targets are the Si and Ge single component systems in the liquid state. To train the machine learning model, Gaussian process regression is performed with the atomic fingerprints which express the local structure around the target atom. The training and test data are generated by the molecular dynamics (MD) based on DFT. We first report the accuracy of ML forces when both test and training data are generated from the DFT-MD simulations at a same temperature. By comparing the accuracy of ML forces at various temperatures, it is found that the accuracy becomes the lowest around the phase boundary between the solid and the liquid states. Furthermore, we investigate the transferability of ML models trained in the liquid state to temperature and pressure. We demonstrate that, if the training is performed at a high temperature and if the volume change is not so large, the transferability of ML forces in the liquid state is high enough, while its transferability to the solid state is very low.

physics.comp-ph↗