TRACER: Trajectory Risk Aggregation for Critical Episodes in Agentic Reasoning
Sina Tayebati, Divake Kumar, Nastaran Darabi, Davide Ettori, Ranganath Krishnan, Amit Ranjan Trivedi · 2026
Read TRACER: Trajectory Risk Aggregation for Critical Episodes in Agentic Reasoning by Sina Tayebati, Divake Kumar, Nastaran Darabi, Davide Ettori, Ranganath Krishnan, Amit Ranjan Trivedi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-11409.txt.
Description
Agent failures often hide in rare key episodes, not in average confidence. Sina Tayebati and five co-authors offer TRACER, a path-level doubt metric for tool-using agent contact with humans. It combines content-aware surprisal with local awareness, repeats signals, and tool-grounded sense gaps.
A tail-focused risk functional with a max-mixed step risk surfaces decisive anomalies that single-shot proxies miss. Evaluated on tau-bench for failure forecast and selective execution, TRACER improves discrimination and selective scores greatly over rivals. The work enables earlier, more accurate finding of doubt in complex conversational tool use, where looping, incoherent tool calls, or miscoordination trigger failure.
Catalog record
| CREATOR | Sina Tayebati, Divake Kumar, Nastaran Darabi, Davide Ettori, Ranganath Krishnan, Amit Ranjan Trivedi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-11409