Open research on AI agent reliability, execution forensics, environments, and post-training.
| Project | Venue / Links | Focus |
|---|---|---|
| TraceTrial-Bench | arXiv:2610.32285 · Report · Leaderboard | 95 mechanism-focused tasks across 7 dimensions (C, E, G, L, O, T, V) separating model capability from agent harness reliability |
| Judging-by-Evidence | Research Code | Typed evidence acquisition and verification for Agent-as-a-Judge |
| MM-VeriRubric | EMNLP 2026 | Verifiable evidence-grounded rubrics for multimodal reward models |
| DPS | EMNLP 2026 | Disentangled privileged states for simulator-assisted multi-turn dialogue policy training |
| Awesome-Agentic-Environments | Curated Survey | Papers, benchmarks, and frameworks for building and scaling LLM agent environments |
Using TraceTrial-Bench as a diagnostic probe, we actively contribute bug reports, security fixes, and runtime improvements back to open-source agent frameworks:
- OpenHands (
software-agent-sdk) — 10 pull requests addressing fail-closed sub-agent confirmation & security analyzer inheritance (#7001,priority:high), secret rotation masking (#7003), cross-restart sub-agent recovery (#6986), MCP tool collision isolation & structured payload preservation (#6988, #6989), sub-agent event streaming (#6997), entity-preserving condensation (#6998), and deterministic terminal timeout/truncation contracts (#6987, #7024, #7025). - OpenCode (
anomalyco/opencode) — Fixes and proposals for direct MCP structured results (#64838), sub-agent session permission inheritance (#64841), non-interactive--format jsonbackground event streaming (#64852), and native interactive PTY sessions (#65017). - Aider (
Aider-AI/aider) — Non-interactive--yes-shell-commandsexecution and shell output reflection (#6002, #6003).
Xiaohongshu
|
WeChat Official Account
|

