Search arXiv⌕ Search

arXiv subjects

Chenlei Guo

Publications and source records attributed to Chenlei Guo.

At least 19 recordsLinked to original sources
arXiv2026

MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning

Agent memory enables large language model (LLM) agents to reuse past experiences without updating their parameters. However, semantic similarity alone does not guarantee memory usefulness: downstream performance depends on how memories are selected and presented to the agent. We propose MemComposer, a memory composer that jointly examines the current query and retrieved candidates to decide which experiences to include, how many to select, and how to order them. We train the composer using Experience Policy Optimization (ExPO), a GRPO-based framework that evaluates assembled memory contexts through complete agent interactions and uses task-success feedback to optimize the assembly policy. Across three $τ^2$-bench domains with two frozen executors, MemComposer outperforms the strongest baseline for each executor while reducing average memory-token usage by 16.9% and 4.9%, respectively. The learned policy also transfers to unseen domains, executors, and memory banks, and benefits from memories generated by stronger models without further training. These results demonstrate the value of learning query-specific memory assembly using downstream task feedback.

cs.AI↗
arXiv2026

A Shell Collision Scenario for EP251222b/GRB 251222A: Connecting Prompt Emission and Double-Peaked Afterglow

GRB 251222A is a long-duration gamma-ray burst at a spectroscopic redshift of $z=3.171$. It was detected by the EP as the soft X-ray transient EP251222b before and during the gamma-ray emission. We performed a multi-wavelength (EP/WXT, fermi/GBM, swift/XRT and optical/NIR) investigation of this event. Its prompt soft X-ray and gamma-ray features, together with two afterglow rebrightenings, allow us to examine a possible connection between the prompt ejecta and the afterglow. We consider a multiple-shell scenario in which a faster, narrower shell catches up with a slower, wider shell. We assume that the selected pulse groups trace different ejecta and that the late X-ray flare marks their collision. Under these assumptions, we represent the post-collision outflow by an effective merged core and an outer annular component. A joint fit to the multi-band afterglow gives a Lorentz factor of $Γ_c\approx293$ for the merged core and $Γ_{1,w}\approx228$ for the outer component. The sample-wise collision mapping gives an energy release of $E_{\rm diss}=(1.6^{+2.1}_{-0.9})\times10^{52}$ erg and a dynamical dissipation efficiency of $0.37^{+0.07}_{-0.14}\%$. EP251222b can therefore be used to test a prompt--afterglow connection through shell collisions, although it does not uniquely establish this interpretation.

astro-ph.HE↗
arXiv2026

Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution

Skills, reusable procedural guidance added at inference, can substantially improve LLM downstream performance (Li et al., 2026). Prior work retrieves skills from a bank by semantic relevance, then uses them as inference-time patches or for model distillation. The individual utility of each skill, however, is largely neglected. We first show that, in on-policy distillation where skill-conditioned policies serve as teachers, fewer than 25% of retrieved skills provide useful distillation signals. We then propose SGUID, a method for selecting a compact subset of skills for distillation. SGUID retains a skill only if it consistently yields effective learning signals during training. The selected skills are then distilled to produce a better model. Our results show that not all skills are worth distilling. Across four models from the Olmo and Qwen families, distilling 6 selected skills matches or exceeds full-bank distillation in mean avg@12 on three of the four models, and on all four after a second round that distills 3 newly selected skills, while the full banks are up to 11x larger. Importantly, SGUID supports stable model-skill co-evolution: after a distillation round, a new candidate bank is curated from the updated model's rollouts, and SGUID selects which skills to internalize next. In the second round, this loop selects 3 new skills and improves Qwen3-8B from 64.3% to 66.3%. The selection step is essential for stability: on Qwen3-4B, naively updating the model with unfiltered skills degrades performance, including a 0.3 percentage point drop on HMMT25, whereas SGUID improves HMMT25 by 0.5 points after the first round and 1.1 points after the second. These results identify skill selection as the key mechanism for stable model-skill co-evolution.

cs.CL↗
arXiv2026

WOVEN: Weaving Visual World Modeling into Multimodal LLMs

Multimodal large language models (MLLMs) struggle with spatial, embodied, physical, and temporal reasoning. We hypothesize that these failures reflect a shared deficit in visual transition reasoning, and test whether this capability can serve as a shared training primitive, one that different models can learn from different supervision sources and reuse across different tasks, with a systematic training recipe. Existing benchmarks document these deficits separately but do not support controlled comparisons across scenes, actions, and reasoning operations. We therefore introduce WOVEN, a training source and benchmark for visual transition reasoning that organizes transition supervision by scene, action, and reasoning type, using diverse, realistic rollouts from video-pretrained generative models: 36,076 examples across 20 scene types, 5 action types, and 8 reasoning types. We first evaluate 38 frontier MLLMs (e.g., GPT-5.4 and Qwen3-VL-235B-A22B) and find a substantial and systematic deficit: even the strongest models fall far below humans, and the failures recur across model families and persist with scale. We then train MLLMs at multiple scales on WOVEN and find that they learn a shared capability that transfers broadly: training subsets of only about 2,000 items each collectively improve 22 of 26 external benchmarks by up to 27.3 percentage points, and WOVEN data can replace 30-50% of a task's own training data with comparable accuracy. Controlled comparisons further yield a training recipe for visual world modeling, validated prospectively on held-out benchmarks: select supervision by the reasoning operation it teaches rather than by the actions, scenes, or domains it shows, and prefer larger changes to the visual state for robustness. Our work establishes visual transition reasoning as a reusable foundation for systematic visual world-model training in MLLMs.

cs.CV↗
arXiv2026

EnSIMem: Entity-Structured Indexing for Long-Term Agent Memory

An agent that interacts with users over long periods must recall facts, preferences, events, and changes from a continuously growing interaction history. Existing memory systems often compress interactions into generic summaries or retrieve anonymous text chunks, making it difficult for an agent to identify the correct entity, property, and supporting evidence. We present EnSIMem, an entity-structured long-term memory architecture for an agent. During offline construction, the system organizes interactions into theme-coherent episodes and builds dialogue-grounded index entries of the form [entity][entity_type][property: value]. Each entry preserves its source turns, temporal information, and available multimodal fields. During online interaction, the agent's request is decomposed into evidence requirements whose properties are aligned with the memory index. Entity-property lookup and adaptive retrieval then collect the evidence needed for point, temporal, compositional, and aggregation reasoning. The agent generates its response from the preserved source evidence rather than from lossy memory summaries. On long-term agent-memory benchmarks, EnSIMem achieves high answer accuracy while maintaining compact, evidence-focused contexts. These results show that entity-structured indexing and episode-level provenance provide a reliable foundation for long-term memory in agents. The code of our model is available at https://github.com/RamonMeng/EnSIMem.

cs.AI↗
arXiv2026

Scaling Automatic Research Agents via World Models

Automating empirical research is a long-standing direction of AI. Recent automatic research (AutoResearch) agents bring this goal within reach, as modern LLMs show the capability to independently implement solutions and learn from the execution outcomes. Behind these gains, post-training (especially RL) plays a central role. In this paper, we identify a fundamental tension when scaling RL for these agents: the two components of every AutoResearch trajectory (agent generation and environment execution) scale in very different manners, since all generation shares compute through batching, while each execution occupies its exclusive sandbox and real machine time. As a result, the environment execution dominates the training cost and becomes the bottleneck as trajectories grow. To resolve this tension, we propose World Model RL (WMRL), which replaces environment execution with a world model to remove this bottleneck. Additionally, the world model can be imperfect, as its rewards are corrupted by bias and noise. Therefore, we further equip WMRL with two mitigations, Online Debiasing and Inverse-Variance Denoising, which offset the bias and suppress the noise respectively. Theoretically, we prove that both mitigations of WMRL strictly improve the convergence guarantee. Empirically, WMRL accelerates training by 3-4x on various tasks at different agent scales, while exceeding the performance of standard RL baselines. Moreover, our post-trained 4B and 9B agents outperform much larger open-weight agents of 48B and 120B on held-out benchmarks. Beyond AutoResearch, WMRL also transfers to post-training embodied VLA policies, which demonstrates the generalizability of our method.

cs.LG↗
arXiv2026

PA3: Policy-Aware Agent Alignment through Chain-of-Thought

Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the 'needle-in-a-haystack' problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel Policy Recall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

cs.CL↗
arXiv2026

Time-dependent Accretion Disks in Tidal Disruption Events: Long-term Light Curves

The time-dependent accretion disk has been applied to explain the light curves observed in tidal disruption events (TDEs). Radiation pressure instability is expected to be an important factor that can shape the evolution of the accretion disk. In this paper, we upgrade the time-dependent disk model in Guo $\&$ Qiao by incorporating an index $μ$ (with the stress tensor $\propto p^μp^{1-μ}_{\rm{gas}}$, where $p=p_{\rm{gas}} + p_{\rm{rad}}$) for the modified viscosity, and a parameter $f_{\rm{w}}$ for the strength of the wind. Meanwhile, we adopt a more realistic fallback rate to inject into the disk. After systematically testing the effects of the newly incorporated parameters, we find that $f_{\rm{w}}$ can affect the time when the instability occurs, while $μ$ can influence the occurrence and variation magnitude of the radiation pressure instability. When $μ<0.4$, the radiation pressure instability is completely removed from the disk, and the light curves evolve stably without large-scale magnitude variation. When $μ>0.4$, the light curves can show oscillations caused by the instability, or drop steeply when the instability occurs and become flat in the late-time evolution, which mainly occurs when the viscous viscosity parameter $α$ is small or the impact parameter $β$ is small. Since the drop magnitude can be modulated by $μ$, we apply these `decay-to-flatten' light curves to some optical/UV observations in TDEs with different magnitudes of decline. Finally, we discuss the potential application of our model to some special TDEs that show oscillations in the light curves.

astro-ph.HE↗
arXiv2026

Multimodal Policy Internalization for Conversational Agents

Modern conversational agents like ChatGPT and Alexa+ rely on predefined policies specifying metadata, response styles, and tool-usage rules. As these LLM-based systems expand to support diverse business and user queries, such policies, often implemented as in-context prompts, are becoming increasingly complex and lengthy, making faithful adherence difficult and imposing large fixed computational costs. With the rise of multimodal agents, policies that govern visual and multimodal behaviors are critical but remain understudied. Prior prompt-compression work mainly shortens task templates and demonstrations, while existing policy-alignment studies focus only on text-based safety rules. We introduce Multimodal Policy Internalization (MPI), a new task that internalizes reasoning-intensive multimodal policies into model parameters, enabling stronger policy-following without including the policy during inference. MPI poses unique data and algorithmic challenges. We build two datasets spanning synthetic and real-world decision-making and tool-using tasks and propose TriMPI, a three-stage training framework. TriMPI first injects policy knowledge via continual pretraining, then performs supervised finetuning, and finally applies PolicyRollout, a GRPO-style reinforcement learning extension that augments rollouts with policy-aware responses for grounded exploration. TriMPI achieves notable gains in end-to-end accuracy, generalization, and robustness to forgetting. As the first work on multimodal policy internalization, we provide datasets, training recipes, and comprehensive evaluations to foster future research. Project page: https://mikewangwzhl.github.io/TriMPI.

cs.CL↗
arXiv2026

Light curves of time-dependent accretion disk in tidal disruption events

Tidal disruption events (TDEs) are believed to be an ideal laboratory for studying the evolution of accretion flow around a supermassive black hole (BH). In general, the mass feeding rate to the BH is suggested to be super-Eddington initially, and evolves to be sub-Eddington on timescales of years. In this paper, we carry out calculations of the time-dependent evolution of accretion disk in the standard environment of TDE, i.e., injecting matter at the circularization radius of the stellar debris in the form of $\dot M_{\rm inject} \propto t^{-5/3}$. One of the main findings is that when $\dot M_{\rm inject}$ evolves to a value around the Eddington accretion rate, the radiation pressure instability occurs. We test the influence of the model parameters on the light curves, such as the BH mass $M_{\rm BH}$, viscosity parameter $α$, and mass-injecting radius $R_{\rm{out}}$, all of which are found to affect the light curves to some extent. In most cases, we find that the light curves oscillate significantly due to the radiation pressure instability. As an exception, when $α$ is small or $R_{\rm{out}}$ is large, we find that the oscillations are completely suppressed. In this case, the light curve drops steeply and then becomes flat in the late-time evolution, which we apply to explain the observed ultraviolet (UV) light curves of ASASSN-15oi and ASASSN-14ae together with the assumption of a photosphere. Finally, we discuss the potential applications of our time-dependent accretion disk model to explaining multi-band light curves of TDEs in the future.

astro-ph.HE↗
arXiv2026

Beyond Perfect APIs: A Comprehensive Evaluation of LLM Agents Under Real-World API Complexity

We introduce WildAGTEval, a benchmark designed to evaluate large language model (LLM) agents' function-calling capabilities under realistic API complexity. Unlike prior work that assumes an idealized API system and disregards real-world factors such as noisy API outputs, WildAGTEval accounts for two dimensions of real-world complexity: 1. API specification, which includes detailed documentation and usage constraints, and 2. API execution, which captures runtime challenges. Consequently, WildAGTEval offers (i) an API system encompassing 60 distinct complexity scenarios that can be composed into approximately 32K test configurations, and (ii) user-agent interactions for evaluating LLM agents on these scenarios. Using WildAGTEval, we systematically assess several advanced LLMs and observe that most scenarios are challenging, with irrelevant information complexity posing the greatest difficulty and reducing the performance of strong LLMs by 27.3%. Furthermore, our qualitative analysis reveals that LLMs occasionally distort user intent merely to claim task completion, critically affecting user satisfaction.

cs.CL↗
arXiv2025

Analyzing and Internalizing Complex Policy Documents for LLM Agents

Large Language Model (LLM)-based agentic systems rely on in-context policy documents encoding diverse business rules. As requirements grow, these documents expand rapidly, causing high computational overhead. This motivates developing internalization methods that embed policy documents into model priors while preserving performance. Prior prompt compression work targets generic prompts, but agentic policy documents span multiple complexity levels and require deeper reasoning, making internalization harder. We introduce CC-Gen, an agentic benchmark generator with Controllable Complexity across four levels, enabling systematic evaluation of agents' ability to handle complexity and offering a unified framework for assessing policy internalization. Our analysis shows that complex policy specifications governing workflows pose major reasoning challenges. Supporting internalization with gold user agent interaction trajectories containing chain-of-thought (CoT) annotations via supervised fine-tuning (SFT) is data-intensive and degrades sharply as policy complexity increases. To mitigate data and reasoning burdens, we propose Category-Aware Policy Continued Pretraining (CAP-CPT). Our automated pipeline parses policy documents to extract key specifications, grouping them into factual, behavioral, and conditional categories, and isolating complex conditions that drive workflow complexity. This guides targeted data synthesis and enables agents to internalize policy information through an autoregressive pretraining loss. Experiments show CAP-CPT improves SFT baselines in all settings, with up to 41% and 22% gains on Qwen-3-32B, achieving 97.3% prompt length reduction on CC-Gen and further enhancing tau-Bench with minimal SFT data.

cs.AI↗
arXiv2025

Early evolution of super-Eddington accretion flow in tidal disruption events

Tidal disruption events (TDEs) are luminous black hole (BH) transient sources, which are detected mainly in X-ray and optical bands. It is generally believed that the X-ray emission in TDEs is produced by an accretion disc formed as the stellar debris accreted onto the central BH. The origin of the optical emission is not determined, but could be explained by the `reprocessing' model with the X-ray emission reprocessed into optical band by a surrounding optically thick envelope or outflow. In this paper, we performed radiation hydrodynamic simulations of super-Eddington accretion flow with Athena++ code in the environment of TDEs, i.e., injecting a continuous mass flow rate at the circularization radius in the form of $\dot M_{\rm inject} \propto t^{-5/3}$ for the mass supply rate. We show that a significant fraction of the matter in the accretion inflow are blowed off forming outflow, and the properties of the outflow are viewing-angle dependent. We further calculate the emergent spectra of such an inflow/outflow system for different viewing angles with the method of Monto Carlo radiative transfer. Based on the emergent spectra, we show that the observed features of TDEs, such as the X-ray and optical luminosities, the blackbody temperature of X-ray and optical emission and the corresponding emission radii, the ratio of X-ray luminosity to optical luminosity, as well as the evolution of these quantities can be explained in the framework of viewing-angle effect of super-Eddington accretion around a BH.

astro-ph.HE↗
arXiv2024

MEND: Meta dEmonstratioN Distillation for Efficient and Effective In-Context Learning

Large Language models (LLMs) have demonstrated impressive in-context learning (ICL) capabilities, where a LLM makes predictions for a given test input together with a few input-output pairs (demonstrations). Nevertheless, the inclusion of demonstrations leads to a quadratic increase in the computational overhead of the self-attention mechanism. Existing solutions attempt to distill lengthy demonstrations into compact vectors. However, they often require task-specific retraining or compromise LLM's in-context learning performance. To mitigate these challenges, we present Meta dEmonstratioN Distillation (MEND), where a language model learns to distill any lengthy demonstrations into vectors without retraining for a new downstream task. We exploit the knowledge distillation to enhance alignment between MEND and LLM, achieving both efficiency and effectiveness simultaneously. MEND is endowed with the meta-knowledge of distilling demonstrations through a two-stage training process, which includes meta-distillation pretraining and fine-tuning. Comprehensive evaluations across seven diverse ICL task partitions using decoder-only (GPT-2) and encoder-decoder (T5) attest to MEND's prowess. It not only matches but often outperforms the Vanilla ICL as well as other state-of-the-art distillation models, while significantly reducing the computational demands. This innovation promises enhanced scalability and efficiency for the practical deployment of large language models

cs.CL↗
arXiv2023

PersonaPKT: Building Personalized Dialogue Agents via Parameter-efficient Knowledge Transfer

Personalized dialogue agents (DAs) powered by large pre-trained language models (PLMs) often rely on explicit persona descriptions to maintain personality consistency. However, such descriptions may not always be available or may pose privacy concerns. To tackle this bottleneck, we introduce PersonaPKT, a lightweight transfer learning approach that can build persona-consistent dialogue models without explicit persona descriptions. By representing each persona as a continuous vector, PersonaPKT learns implicit persona-specific features directly from a small number of dialogue samples produced by the same persona, adding less than 0.1% trainable parameters for each persona on top of the PLM backbone. Empirical results demonstrate that PersonaPKT effectively builds personalized DAs with high storage efficiency, outperforming various baselines in terms of persona consistency while maintaining good response generation quality. In addition, it enhances privacy protection by avoiding explicit persona descriptions. Overall, PersonaPKT is an effective solution for creating personalized DAs that respect user privacy.

cs.CL↗
arXiv2023

CLICKER: Attention-Based Cross-Lingual Commonsense Knowledge Transfer

Recent advances in cross-lingual commonsense reasoning (CSR) are facilitated by the development of multilingual pre-trained models (mPTMs). While mPTMs show the potential to encode commonsense knowledge for different languages, transferring commonsense knowledge learned in large-scale English corpus to other languages is challenging. To address this problem, we propose the attention-based Cross-LIngual Commonsense Knowledge transfER (CLICKER) framework, which minimizes the performance gaps between English and non-English languages in commonsense question-answering tasks. CLICKER effectively improves commonsense reasoning for non-English languages by differentiating non-commonsense knowledge from commonsense knowledge. Experimental results on public benchmarks demonstrate that CLICKER achieves remarkable improvements in the cross-lingual CSR task for languages other than English.

cs.CL↗
arXiv2023

KG-ECO: Knowledge Graph Enhanced Entity Correction for Query Rewriting

Query Rewriting (QR) plays a critical role in large-scale dialogue systems for reducing frictions. When there is an entity error, it imposes extra challenges for a dialogue system to produce satisfactory responses. In this work, we propose KG-ECO: Knowledge Graph enhanced Entity COrrection for query rewriting, an entity correction system with corrupt entity span detection and entity retrieval/re-ranking functionalities. To boost the model performance, we incorporate Knowledge Graph (KG) to provide entity structural information (neighboring entities encoded by graph neural networks) and textual information (KG entity descriptions encoded by RoBERTa). Experimental results show that our approach yields a clear performance gain over two baselines: utterance level QR and entity correction without utilizing KG information. The proposed system is particularly effective for few-shot learning cases where target entities are rarely seen in training or there is a KG relation between the target entity and other contextual entities in the query.

cs.CL↗
arXiv2022

PENTATRON: PErsonalized coNText-Aware Transformer for Retrieval-based cOnversational uNderstanding

Conversational understanding is an integral part of modern intelligent devices. In a large fraction of the global traffic from customers using smart digital assistants, frictions in dialogues may be attributed to incorrect understanding of the entities in a customer's query due to factors including ambiguous mentions, mispronunciation, background noise and faulty on-device signal processing. Such errors are compounded by two common deficiencies from intelligent devices namely, (1) the device not being tailored to individual customers, and (2) the device responses being unaware of the context in the conversation session. Viewing this problem via the lens of retrieval-based search engines, we build and evaluate a scalable entity correction system, PENTATRON. The system leverages a parametric transformer-based language model to learn patterns from in-session customer-device interactions coupled with a non-parametric personalized entity index to compute the correct query, which aids downstream components in reasoning about the best response. In addition to establishing baselines and demonstrating the value of personalized and context-aware systems, we use multitasking to learn the domain of the correct entity. We also investigate the utility of language model prompts. Through extensive experiments, we show a significant upward movement of the key metric (Exact Match) by up to 500.97% (relative to the baseline).

cs.LG↗