Autoresearch All-Topics Update

State

  • 当前 agenda 的 primary direction 仍是 Agent-Facing Environment Runtime;GUI Grounding Robustness 是 secondary high,RL-based GUI Agent Training 仍被 credit-assignment 拥挤度约束。
  • 本轮按 all-topics recent-window pulse 执行:fetch_and_score.py --date 2026-07-03 --days 3 合并 HF、OpenAlex、CVF 后得到 54 篇候选。arXiv API 返回 429,Lexmount key 未设置,但 HF/OpenAlex/CVF 已足够支撑分流。
  • 上轮遗留队列有 6 篇;本轮消化最高优先级的 Dockerless,新增 AgenticSTS,其余新信号进入 queue/watchlist。

New Signal

1. AFE verifier 不必只能是 hidden oracle

2606-Dockerless 的核心模式是:verifier 先提出少量诊断问题,再让 read-only sub-agents 进入环境找证据,最后由 judge 聚合 verdict。它在 SWE 里用代码文件和 reference patch 工作,但对 AFE 更重要的是抽象:

  • verifier 输入不该只是最终状态或一行 success label;
  • verifier 可以要求环境提供可审计 evidence object;
  • reward / SFT filter 可以来自 evidence-grounded verifier,而不是必须跑昂贵的 hidden test suite。

这提示 AFE-MiniSuite 应把 verifier 设计成三层:agent-safe observations、read-only verifier evidence、hidden oracle label。实验应测“evidence verifier 是否接近 oracle,同时比 prompt-only feedback 更能提升 recovery / stop / avoid-false-completion”。

2. Memory/context 是 runtime contract,不是越长越好

2607-AgenticSTS 对 AFE 的贡献不在 Slay the Spire 2 本身,而在 “每步 prompt 由 typed layers 组成” 这个 contract:

  • L1/L2 是 protocol/schema;
  • L3 是环境规则;
  • L4 是 episodic summaries;
  • L5 是 triggered skills;
  • raw transcript 不跨决策无限追加。

GUI/CUA runtime 也需要类似分层:visible UI tree / screenshot evidence、recent failed action summary、retrieved app manual、task-progress verifier notes、stop/recovery skills。否则“给 agent 更多历史”既不可控,也难以判断收益来自哪一种证据。

3. 全域候选正在分化为三条后续线索

  1. Verifier / reward-source:Dockerless、QVal、PolicyGuard、Decodable Is Not Grounded。这里最贴近 AFE 的是 evidence object 和 verifier feedback。
  2. Long-horizon memory / skill / horizon scaling:AgenticSTS、Agents-A1、Managing Procedural Memory。这条线提醒 AFE 需要 bounded context 与可审计 skill/memory 更新,而不是无限轨迹日志。
  3. VLA / world-action / embodied transfer:ABot-M0.5、Domain Arithmetic、Does VLA Even Know the Basics、Learning to Move Before Learning to Do。这些对 VLA-Survey 有价值,但不应压过当前 primary AFE。

Collected Papers

StatusPaperNote / Reason
digestedDockerless: Environment-Free Program Verifier for Coding Agents2606-Dockerless
digestedAgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents2607-AgenticSTS
queuedScaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agenthorizon scaling / long trajectory infrastructure
queuedDoes VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in VLA ModelsVLA knowledge retention
queuedManaging Procedural Memory in LLM Agentsmemory / skill benchmark
queuedQVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agentsdense reward/verifier evaluation
queued-lowABot-M0.5, Domain Arithmetic, Learning to Move Before Learning to DoVLA/world-action, lower current priority
skipped duplicateQwen-AgentWorld, OpenRath, LUMOS, PersonaVLM, GraphVLM, ActiveVLA, CoA-VLA, MoMa-Kitchen, FedVLAexisting notes / recent reports

Topic Updates

Agent-Facing Environment Runtime

Dockerless + AgenticSTS jointly push AFE from “state API” toward “bounded evidence contract”。一个 AFE observation 不应只是 DOM/UIA dump,而应回答:本步 decision 允许 agent 看到哪些 typed evidence?哪些 evidence 只能 verifier 看?哪些 evidence 会写入 long-horizon memory?哪些 recovery/stop skill 会被触发?

GUI Grounding Robustness

本轮没有新增 grounding-only 精读,但 Decodable Is Not Grounded 仍是下一轮重要候选。它的 blank-image arbiter 与 EvidenceDependence-GUIGrounding 同向,可用于区分“可线性解码”与“真实视觉依赖”。

RL-based GUI Agent Training

TRIAGEQVal 都指向 dense/process reward,但 credit assignment 子方向此前已因拥挤被降级。本轮判断:只保留 QVal 作为 verifier-evaluation 视角,暂不恢复 TRIAGE 深读,除非 Supervisor 明确要重启 credit assignment。

VLA / Embodied / World Model

Does VLA Even Know the Basics? 比 ABot-M0.5 更值得先读,因为它不是又一个 VLA architecture,而是诊断 VLA fine-tuning 后 commonsense/world knowledge retention 的协议。对 VLA survey 来说,这更可能提供跨模型适用的 evaluation insight。

Next Action

下一轮优先级建议:

  1. Decodable Is Not Grounded:补强 GUI/VLM evidence-dependence 诊断。
  2. QVal:判断 dense supervision signal 如何被 cheaply evaluated,服务 AFE verifier/reward design。
  3. Does VLA Even Know the Basics?:更新 VLA survey 的 evaluation/retention 线索。

不要急着读 Agents-A1,除非需要专门整理 long-horizon trajectory scaling。它可能更像系统工程 scaling report,而不是当前 AFE 机制缺口。