Agentic Abstention: Do Agents Know When to Stop Instead of Act?

🕒 Published (v1): 2026-06-27 04:49 UTC · Source: Arxiv · link

Why this paper was selected

Novel abstention framing for agent eval; critical for building reliable harnesses

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

LLM agents lack the ability to know when to stop acting on infeasible tasks. This paper formalizes "Agentic Abstention" as a sequential POMDP decision problem, evaluates 13 LLMs + 2 scaffolds across 28K+ tasks in web shopping, terminal, and QA settings, and finds that timely abstention (abstaining at the earliest warranted step) is very poor across the board. It introduces CONVOLVE, a context-engineering method that distills trajectories into reusable stopping rules, boosting Llama-3.3-70B's timely recall from 26.7% to 57.4% on WebShop.

Problem

Standard LLM abstention treats it as a single-turn answer-or-abstain decision. In agentic settings, the action space is richer (search, click, interact), context evolves over multiple turns, and infeasibility may only become apparent after interacting with the environment. Existing agent evaluations focus on successful task completion, leaving a gap in understanding when agents should stop instead of continuing to act.

Method

• Formalizes agentic abstention as a POMDP with action space \(A = \{\text{ANSWER}, \text{ABSTAIN}, \text{ACT}\}\), where the agent receives observations over turns and must decide when to stop.
• Constructs a benchmark (28,000+ instructions) by adapting three scenarios:
• WebShop (500 solvable + 500 abstention-warranted, split into Request-based and Environment-based abstention)
• Terminal-Bench 2.0 (89 solvable + 188 abstention-warranted)
• AbstentionBench QA (16 datasets, 27,073 samples, converted to sequential setting with Wikipedia search)
• Evaluates 13 LLMs (GPT-5.4-mini, Grok 4.1 Fast, Llama-3.3-70B, Qwen-3 family, Gemma-4-31B-it, GLM-5.1, etc.) and 2 scaffolds (Terminus 2, Codex CLI).
• CONVOLVE: a context-engineering method that runs agent rollouts, uses a reflection model to analyze trajectories, and a curator model to distill them into a structured "playbook" of reusable stopping rules appended to agent context (no parameter updates).

Key Contributions

• Problem formulation: Agentic Abstention as a sequential decision problem with a richer action space than prior single-turn LLM abstention work.
• Large benchmark: 28,000+ instructions across web shopping, terminal, and interactive QA, with both Request-based and Environment-based abstention categories.
• Comprehensive evaluation: 13 LLMs + 2 scaffolds; reveals that timely abstention is persistently low (all models <40% average timely recall) and that larger/reasoning models do not always abstain better.
• CONVOLVE: a parameter-free context engineering method that substantially improves timely abstention by distilling trajectories into a reusable playbook.

Results

• Web shopping: Strongest baseline (GPT-5.4-mini) achieves only 26.7% timely recall (AbsRec@1) and 83.2% overall recall (AbsRec@10). 6/8 models have <0.5 abstention recall after 10 turns.
• Terminal: Best configuration reaches only 21.6% timely recall on abstention tasks. Scaffold choice (Terminus 2 vs. Codex CLI) substantially affects performance.
• QA: All models perform at or below 50% on false premise and underspecified intent categories.
• Scale/reasoning: Larger model scale does not improve timely abstention; reasoning can improve timely recall but reduces overall abstention recall.
• CONVOLVE (on WebShop): Llama-3.3-70B timely recall 26.7% → 57.4%, overall recall 83.2% → 100.0% , without any parameter updates.

Limitations

• Only three environments are tested (web shopping, terminal, QA); generalization to other agentic settings (e.g., web browsing, code generation, robotics) is unmeasured.
• CONVOLVE main-text evaluation is limited to WebShop (terminal and QA results deferred to appendix).
• The QA setting uses a static Wikipedia dump (enwiki-20260101) instead of live search, which may affect ecological validity.
• 10-turn interaction budget is somewhat arbitrary and may not reflect real-world constraints.
• Only two agent scaffolds are compared (Terminus 2 and Codex CLI), and comprehensive model coverage varies across scenarios (exhaustive only on WebShop).

Relevance to Harnesses / Meta-Harnesses

This paper is directly relevant to harness/meta-harness design because it formalizes the scaffold-level decision problem of when to stop acting — a dimension current harnesses (e.g., LangGraph, CrewAI, ADK) largely ignore, defaulting to "try until success, timeout, or max steps." The POMDP framing and CONVOLVE method show that stopping rules can be learned and injected as reusable context engineering, not just hardcoded thresholds. For meta-harness builders, this opens a concrete integration point: a harness could expose an abstention policy as a pluggable module (rule-based, distilled, or learned) that the orchestrator consults after each step, reducing cost and failure cascades on infeasible branches. The weak baseline performance (26.7% to 57.4% with the best method) also signals that current scaffold-level designs silently waste budget on unachievable tasks — a directly actionable gap.