Blog / 博客Blog RSS
关于视觉证据与 agentic AI systems 的研究笔记：多模态评测、科研任务环境、网页 Agent、长程执行、AutoResearch、RSI 与可验证的 harness 工程。 Research notes on visual evidence and agentic AI systems: multimodal evaluation, scientific-task environments, web agents, long-horizon execution, AutoResearch, RSI, and verifiable harness engineering.
Paper Introductions
Choose a paper or watch its introduction. English narration with English, Simplified Chinese and bilingual captions.
-
Persona Dosing: Calibrated Activation Steering for Graded Trait Control
-
Self-Evolving Agents via Likelihood-Guided Tool-Space Optimization
-
xDailyBench: Benchmarking LLMs on Professional Consultation for Real-Life Problems
-
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
-
Dr. Claw: An AI Scientist Workspace for Vibe Research
-
Aspire: Can Models Self-Evolve from Vague Goals?
-
S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?
-
WebWorld: The Browser as a World Model for Self-Improving Web Code
-
VGI-Bench: Probing Visual Intelligence in Video Generation Models
-
StartupBench: Benchmarking General-Purpose Agents on Market-Validated End-to-End Workflows
-
MedClaw: Heuristic Agent Harness for Long-Horizon Surgical Video Reasoning
-
Enhancing Vector Quantization with Distributional Matching: A Theoretical and Empirical Study
-
Distributional Matching for Vector Quantization: A Unified Theoretical and Empirical Framework
-
Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models
-
LP-SFT: Keeping Plausible Alternatives Alive in Supervised Fine-Tuning
-
Learning from the Self-future: On-policy Self-distillation for dLLMs
-
CompRank: Efficient LLM Reranking via Token-Level Compression and Decoding-Free Scoring
-
Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic Tasks in Real-World Professional Fields
-
OpenSkill: Open-World Self-Evolution for LLM Agents
-
Where, What, Why, and Importance: Structured Defect Grounding for Text-to-Image Feedback
-
MIRA: Mid-training Rubric Anchoring for Source-Aware Data Selection
-
RewardHarness: Learning Human Preferences for Image Editing with Only 100 Demonstrations
-
ClawBench: Can AI Agents Complete Everyday Online Tasks?
-
Watch Before You Answer: Learning from Visually Grounded Post-Training
-
Dr. Bench: A Multidimensional Evaluation for Deep Research Agents, from Answers to Reports
-
VideoScore2: Think before You Score in Generative Video Evaluation
-
IntentionReasoner: Facilitating Adaptive LLM Safeguards through Intent Reasoning and Selective Query Refinement
-
WikiGap: Promoting Epistemic Equity by Surfacing Knowledge Gaps Between English Wikipedia and other Language Editions
-
StructEval: Benchmarking LLMs’ Capabilities to Generate Structural Outputs
个人智能体需要一份成果契约Personal Agents Need an Outcome Contract
Dots、Grok Bot、Claude自托管环境与Muse：暂停、共享登录、状态恢复和成果验证的真实边界。Dots, Grok Bot, Claude self-hosting and Muse: stop scope, shared credentials, state recovery and proof of outcome.
Claude Opus 5.5：分数背后，究竟是谁在回答？Claude Opus 5.5: The Model Behind the Score
54.61% 与 0/2,872 如何同时成立？追踪安全回退、攻击分母与真实执行路径。How 54.61% and 0/2,872 coexist. Follow safety fallbacks, attack denominators and the actual execution path.
GPT-6.1 Sol 安全卡：评估智能体，先看分母GPT-6.1 Sol's Safety Card: Read the Denominator
28/49,650与23.5%回答不同问题：内部模拟、警告控制、局部回退和行动监控的证据边界。28/49,650 and 23.5% answer different questions. What internal simulation, warning tests and action-aware monitors establish.
答案看起来对，证据真的对吗？三个研究问题When an AI Answer Looks Right: Three Research Questions About Evidence
从结构化输出、引用检索和视频理解出发，检查 AI 答案实际使用的证据。A reading path from structural outputs and citation retrieval to visual dependence in video answers.
检索到了论文，下一步怎么核查主张？You Found the Paper. Does It Support the Claim?
三个手写主张、三种处理决定，以及可下载的证据核查工作表。Three worked claim checks and downloadable evidence-review worksheets.
字段齐全，数值却错了All the Fields Are There. The Value Is Still Wrong.
四个可复跑的手写案例，区分语法解析、字段路径和业务值检查。Four runnable teaching cases separate parsing, key paths and application values.
看起来合理，不等于视觉证据成立What Counts as Visual Evidence? Five Questions for Images and Video
从 VidGround、VideoScore2、VGI-Bench、SDG 与 MedClaw 出发，区分文本捷径、视频评分、过程正确性、局部缺陷和跨时间证据。Five studies, five different checks: text shortcuts, video scores, valid processes, localized defects, and evidence across time.
Agent 自我改进之后，真的变好了吗？Did the Agent Actually Improve? Three Tests for Self-Development
从学习目标、交互经验与执行框架三个角度，检查改进是否迁移到未见任务；附三篇公开预印本的阅读路线与局限。A reading guide to three public preprints: when changes to learning goals, experience, and agent infrastructure transfer beyond development feedback—and when they do not.
知识地图 / 推荐阅读顺序Knowledge Map / Recommended Reading Order
- Agent Research Environments：先问任务对什么工件、何种反馈和 verifier 负责，再把逐步拦截、提交、证据、reward 和最终接纳拆开。Agent research environments: first ask what artifact, feedback, and verifier a task holds an agent accountable for, then separate step interception, commitment, evidence, reward, and final acceptance.
- ClawBench / 网页 Agent：把 self-host 网站、Browser / Computer Use、受控提交拦截、证据轨与独立 verifier 合成一个可训练、可反驳的 RL 设计。ClawBench / web agents: combine self-hosted websites, Browser / Computer Use, controlled submission interception, evidence trails, and independent verifiers into a trainable, falsifiable RL design.
- 长程 Agent：把 long horizon 定义为可恢复的状态机——什么必须 checkpoint，怎样记录 active/wait/recovery 时间，如何在故障和安全退化下仍验证最终完成。Long-horizon agents: define long horizon as a recoverable state machine—what must be checkpointed, how to record active/wait/recovery time, and how completion stays verified under faults and safety degradation.
- AutoResearch：看开发搜索何时只是本地优化，以及候选冻结、独立接纳与复验怎样把结果变成可反驳的研究证据。AutoResearch: learn when development search is only local optimization, and how candidate freezing, independent acceptance, and reruns create falsifiable research evidence.
- RSI：用改进算子、选择、迁移与完整性四条曲线，检验一条 lineage 是否真的从受限优化升级到“改进改进者”。RSI: use improvement-operator, selection, transfer, and integrity curves to test whether a lineage really moves from bounded optimization to “improving the improver.”
- Harness 架构：真实 coding agents 怎样组织 loop、tools、skills、budget、sandbox 与训练轨迹。Harness architecture: how real coding agents organize loops, tools, skills, budgets, sandboxes, and training traces.
- LLM RL 基础：PPO / DPO / GRPO、reward、critic 与 verifiable reward。LLM RL foundations: PPO / DPO / GRPO, reward, critic, and verifiable rewards.
Agent Research Environments：到底在测什么Agent Research Environments: What Is Actually Being Measured?
一张测量地图：MLS-Bench、AutoLab、EdgeBench、Terminal-Bench、SForge、Harbor、Browser Use 与 Computer Use 分别在测什么，以及怎样把逐步拦截、提交、证据、reward 和最终接纳拆开。 A measurement map for MLS-Bench, AutoLab, EdgeBench, Terminal-Bench, SForge, Harbor, Browser Use, and Computer Use, including how to separate step interception, commitment, evidence, reward, and final acceptance.
网页 Agent 环境：把 Browser Use 变成可验证的 RLWeb-Agent Environments: Making Browser Use Verifiable for RL
一份设计说明：自托管站点、DOM/CDP 与 computer-use 接口、受控提交拦截、独立 evaluator，以及 PPO / GRPO 应如何被公平地比较。 A design note for self-hosted sites, DOM/CDP and computer-use interfaces, controlled submission interception, independent evaluators, and a fair PPO / GRPO comparison.
长程 Agent：状态、恢复、计划与可验证的数小时执行Long-Horizon Agents: State, Recovery, Planning, and Verifiable Hours-Long Execution
先区分 long task、单题大工程与科学 workflow，再将 long horizon 定义为可恢复的状态机：什么必须 checkpoint，怎样记录 active/wait/recovery/human 时间，如何在故障和安全退化下仍验证最终完成。 First separate long tasks, single-project challenges, and scientific workflows, then define long horizon as a recoverable state machine: what must be checkpointed, how to record active/wait/recovery/human time, and how completion remains verified under faults and safety degradation.
AutoResearch：把研究循环变成可证伪的系统AutoResearch: Turning a Research Loop into a Falsifiable System
AutoResearch 是方法论，不是里程碑。开发搜索、候选冻结、独立接纳与干净复验将本地刷分、可信研究循环和 RSI 分开。 AutoResearch is a methodology, not a milestone. Development search, candidate freezing, independent acceptance, and clean reruns separate local score chasing, credible research loops, and RSI.
RSI：条件性的理论，不是已达成的里程碑RSI: A Conditional Theory, Not an Achieved Milestone
从 Gödel Machine 到受限自我优化：用改进算子、选择、迁移与完整性四条曲线，检验一条 lineage 是否真的支持强 RSI。 From the Gödel Machine to bounded self-optimization: use improvement-operator, selection, transfer, and integrity curves to test whether one lineage supports strong RSI.
6 个 coding agent 到底怎么跑：harness 架构对比How 6 Coding Agents Actually Run: A Harness Architecture Comparison
深度拆解 hermes-agent、claw-code、codex、opencode、openclaw 和 pi：一次 turn 的控制流、Tools / Skills / Harness 分层，以及为什么 training-friendliness 是好的 harness design criterion。 A deep dive into hermes-agent, claw-code, codex, opencode, openclaw, and pi: one-turn control flow, the Tools / Skills / Harness stack, and why training-friendliness is a useful criterion for good harness design.
LLM RL 核心算法：PPO、DPO、GRPOCore LLM RL Algorithms: PPO, DPO, and GRPO
一张紧凑地图：PPO 对应经典 online RLHF，DPO 对应离线偏好优化，GRPO 对应可验证推理任务里的 critic-free RL。 A compact map: PPO for classic online RLHF, DPO for offline preference optimization, and GRPO for critic-free RL on verifiable reasoning tasks.