15 papers · 1 filter
Beyond Distribution Matching: Selective Margin Preservation for Supervised Fine-Tuning
Yuxuan Wang, Guangxiang Zhao, Xinyu Tan +6
Supervised fine-tuning (SFT) changes logit margins among non-target tokens at teacher-forced prefixes. Matching a local reference distribution keeps those probabilities and also pe…
ScienceArena: Benchmarking LLMs on Latest Scientific Olympiad Competitions
Guangxiang Zhao, Qilong Shi, Xusen Xiao +13
Benchmark saturation and data contamination increasingly obscure genuine scientific reasoning in frontier LLMs. We introduce \textsc{ScienceArena}, an olympiad-style benchmark from…
ArborMem: Navigating Interaction States with Memory Forests
Zongwei Lv, Yuemeng Xu, Yilun Yao +8
Large language models increasingly serve as persistent conversational assistants, requiring memory that preserves relevant experience and maintains continuity across interactions.…
When Do Prompt-Side Agent Playbooks Transfer? Accuracy, Cost, and Runtime Shift in Agent Deployment
Weihong Lin, Lin Sun, Xiangzheng Zhang
Prompt-side playbooks can improve tool-using language agents without retraining, but their portability beyond the source setting is unclear. We study frozen playbook transfer under…
External Experience Serving in Production LLM Systems: A Deployment-Oriented Study of Quality-Cost Trade-offs
Lin Sun, Heming Zhang, Xiangzheng Zhang
Production LLM systems accumulate reusable operational experience, but the practical deployment issue is not merely whether such experience can help. It is how different serving st…
RealClawBench: Live OpenClaw Benchmarks from Real Developer-Agent Sessions
Zongwei Lv, Yaoming Li, Zhewen Tan +7
Agent benchmarks should reflect what users actually ask deployed agents to do, yet existing benchmarks often miss key realism properties of real developer-agent sessions. We introd…
A Primer in Post-Training Reasoning Data: What We Know About How It Works
Yaoming Li, Guangxiang Zhao, Qilong Shi +3
Post-training has become a primary driver of recent progress in large reasoning models, and reasoning data are often the key variable determining whether this stage succeeds. Work…
Harness-Bench: Evaluating Agent Harnesses across Models and Realistic Workflows
Yilun Yao, Xinyu Tan, Chao-Hsuan Liu +10
LLM agents are increasingly deployed as executable systems that use tools, modify workspaces, and produce concrete artifacts. In such workflows, performance depends not only on the…
MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection
Zhewen Tan, Yilun Yao, Huiyan Jin +9
Large language model agents increasingly rely on persistent memory to store past interactions, retrieve relevant demonstrations, and improve long-horizon task execution. However, t…
When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation
Lin Sun, Wang Dexian, Jingang Huang +4
Industrial Retrieval-Augmented Generation (RAG) systems depend on optical character recognition (OCR) to transform visual documents into text. Existing OCR benchmarks rely on chara…
Thinking with Reasoning Skills: Fewer Tokens, More Accuracy
Guangxiang Zhao, Qilong Shi, Xusen Xiao +3
Reasoning LLMs often spend substantial tokens on long intermediate reasoning traces (e.g., chain-of-thought) when solving new problems. We propose to summarize and store reusable r…
Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging
Weihong Lin, Lin Sun, Qilong Shi +6
Model merging has emerged as a promising paradigm for composing the capabilities of large language models by directly operating in weight space, enabling the integration of special…
Beyond Static Alignment: Hierarchical Policy Control for LLM Safety via Risk-Aware Chain-of-Thought
Jianfeng Si, Lin Sun, Weihong Lin +1
Large Language Models (LLMs) face a fundamental safety-helpfulness trade-off due to static, one-size-fits-all safety policies that lack runtime controllabilityxf, making it difficu…
TriPlay-RL: Tri-Role Self-Play Reinforcement Learning for LLM Safety Alignment
Zhewen Tan, Wenhan Yu, Jianfeng Si +9
In recent years, safety risks associated with large language models have become increasingly prominent, highlighting the urgent need to mitigate the generation of toxic and harmful…
BEAR: Budgeted Evidence Allocation for Multi-Document Reasoning
Lin Sun, Linglin Zhang, Jingang Huang +3
We argue that multi-document reasoning is constrained not only by how much text a model can read, but also by how limited query-time evidence budget is allocated across documents a…