What it is
Defines agentic abstention as a sequential stop-or-act decision for tool-using LLM agents, benchmarks it across 28k tasks, and improves it with context-engineered stopping rules. Instead of single-turn answer-or-abstain, treats abstention as a sequential agent decision over tool use. Adds CONVOLVE, which distills full trajectories into reusable stopping rules appended to context, avoiding parameter updates.
Key results
- WebShop: timely recall 57.4 vs 26.7 for Llama-3.3-70B with CONVOLVE
- WebShop: overall recall 100.0 vs 83.2 for Llama-3.3-70B with CONVOLVE
- Across tested agents, average timely abstention recall stays below 40%
Shows that the hard part is timely stopping after evidence accumulates, and that context-level stopping rules can fix much of this without retraining.