Taming Actor-Observer Asymmetry in Agents
via Dialectical Alignment

Bobo Li¹  Rui Wu²  Zibo Ji³  Meishan Zhang⁴  Hao Fei^5*
Min Zhang⁴  Mong-Li Lee¹  Wynne Hsu¹

¹National University of Singapore   ²Sichuan University   ³University of Minnesota Twin Cities
⁴Harbin Institute of Technology, Shenzhen   ⁵University of Oxford

---

Overview

LLM agents exhibit Actor-Observer Asymmetry (AOA): as actors they blame external factors, as observers they blame internal faults. ReTAS (Reasoning via Thesis-Antithesis-Synthesis) mitigates this via dialectical SFT + GRPO.

Setup

curl -LsSf https://astral.sh/uv/install.sh | sh   # if uv not installed

git clone https://github.com/unikcc/ReTAS.git && cd ReTAS
uv venv --python 3.10
source .venv/bin/activate

uv pip install -e ".[train]"         # SFT + GRPO training + AFB + SalesArena
uv pip install -e ".[train,serve]"   # also vLLM for local serving

requires-python = ">=3.10". Blackwell / CUDA 13: public vLLM wheels are cu12 and will downgrade torch to CPU, so skip .[serve] and serve via SGLang nightly cu13 or NGC instead.

Data

Dataset lives on HuggingFace (gated, auto-approved).

1. Visit the page above, click Agree and access.
2. hf auth login with a Read token from https://huggingface.co/settings/tokens.
3. bash scripts/download_data.sh to fetch and lay out FinQA + Spider under the paths the trainers expect.

Modules

Each directory has its own README with training / evaluation commands:

Path	Description
FinQA/	TAS-SFT + GRPO on FinQA (financial reasoning)
Spider/	TAS-SFT + GRPO on Spider (text-to-SQL, 166 SQLite DBs)
misc/AFB/	Ambiguous Failure Benchmark, 10 domains × 2 scenarios
misc/SalesArena/	Multi-agent negotiation under 4 review mechanisms

Results

Citation

@inproceedings{li2026taming,
  title={Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment},
  author={Li, Bobo and Wu, Rui and Ji, Zibo and Zhang, Meishan and Fei, Hao and Zhang, Min and Lee, Mong-Li and Hsu, Wynne},
  booktitle={Proceedings of the Annual Meeting of the Association for Computational Linguistics},
  year={2026}
}

License

MIT.