Sign InOpen Brain
arXivPaperNeeds Review

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

onPanda turns the first bad token into an annotation boundary, regenerating from each correction to create on-policy supervision with less manual rewriting.

arXiv · Sep 21, 2026
Open Source Open MarkdownOpen JSON
Source Summary

onPanda asks an annotator to find the **first inappropriate token**, replace it from candidates or free-form text, discard the remaining output and regenerate from the corrected prefix. A small controlled study reports **52% lower median annotation time** than manual post-editing.

Practical Implication

For agent tuning, capture corrections at the moment a trajectory first diverges instead of rewriting the final answer wholesale. The resulting positions and paired alternatives can support on-policy SFT and preference data, including tool-connected trajectories.

Agent-Ready Context
onPanda asks an annotator to find the **first inappropriate token**, replace it from candidates or free-form text, discard the remaining output and regenerate from the corrected prefix. A small controlled study reports **52% lower median annotation time** than manual post-editing.

For agent tuning, capture corrections at the moment a trajectory first diverges instead of rewriting the final answer wholesale. The resulting positions and paired alternatives can support on-policy SFT and preference data, including tool-connected trajectories.

The time result comes from a small study, and the supplied material gives no breakdown by task or annotator. Whether the workflow scales to long coding trajectories without adding review overhead remains open.
Connected Context · Feed7 Judgment

This turns the prior call to mine real agent trajectories into a specific annotation primitive: intervene at the first bad token and preserve the corrected prefix as on-policy training evidence. It may make expert review cheaper than full rewrites, but the small timing study does not establish that first-token correction remains efficient or sufficient for long, branching tool-use failures.

Improving Agents is a Data Mining Problem — Vivek Trivedy, LangChainTrace mining identifies recurring production failures; onPanda supplies a concrete way to annotate the first divergence in those traces for subsequent training.State of Data — Sean Cai, Independent / State of DataIt operationalizes the preference for real workflow trajectories by extracting correction data from generated paths rather than relying only on manufactured tasks or saved final answers.Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AIToken-level intervention could reduce the cost of expert labeling in vertical workflows, but it does not remove the candidate’s requirement for domain experts when usefulness lacks an objective answer key.TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning DevelopmentTraceML’s long-horizon planning failures suggest an important boundary: correcting the first inappropriate token may capture a local divergence without representing broader deficiencies in effort allocation or plan revision.
Context Map
benchmarkdata#agent-evals#harness-engineering
Uncertainty
The time result comes from a small study, and the supplied material gives no breakdown by task or annotator. Whether the workflow scales to long coding trajectories without adding review overhead remains open.