PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Han Luo, Bingbing Wen, Lucy Lu Wang

arXiv:2606.28733v1cs.AI

TL;DR

The paper studies when multi-turn agents should stop rather than continue acting when tasks are uncertain or infeasible. It formalizes Agentic Abstention, evaluates agents across more than 28,000 tasks, and introduces CONVOLVE, which improves timely abstention without parameter updates. Agents generally abstain too late or not at all, while CONVOLVE raises Llama-3.3-70B’s WebShop timely recall from 26.7 to 57.4.

  • Problem

    The paper addresses the limited study of when tool-using agents should stop acting, rather than treating abstention as a single-turn answer-or-abstain decision.

  • Method

    The paper formulates agentic abstention as a sequential partially observable decision problem and evaluates 13 LLM-as-agent systems and 2 scaffolds across web, terminal, and QA tasks.

  • Results

    Agents usually abstain too late or not at all across web, terminal, and interactive QA, while CONVOLVE raises Llama-3.3-70B’s WebShop timely recall from 26.7% to 57.4%.

  • Takeaways & Limitations

    Reliable agents require timely stopping behavior in addition to task-completion ability, and reusable stopping rules can improve timely abstention without updating model parameters.

  • Takeaways & Limitations

    The formulation treats clarification requests as abstention rather than as a separate primary action, and episodes terminate when clarification is issued.

Abstract

from arXiv · show

LLM agents are expected to act over multiple turns, using search, browsing interfaces, and terminal tools to complete user goals. Yet not every goal is well specified or achievable in the available environment. In such cases, a reliable agent should recognize that further interaction is unlikely to help and abstain from additional tool calls. We define Agentic Abstention, the problem of deciding when an agent should stop acting under uncertainty. Unlike standard LLM abstention, which is usually evaluated as a single-turn answer-or-abstain decision, agentic abstention is a sequential decision problem: an agent can answer, abstain, or gather more information at each turn, and the need to abstain may only become clear after interacting with the environment. We study this problem across web shopping, terminal environments, and question answering, evaluating 13 LLM-as-agent systems and 2 agent scaffolds on more than 28,000 tasks. Our results show that the main challenge is not only whether agents can abstain, but also when they abstain. Some agents never abstain when they should, while others do so only after many unnecessary interactions. This gap is especially large on tasks where the instruction appears feasible until the environment reveals otherwise (e.g., no valid result matches the instruction). We further find that model scale, reasoning, and agent scaffolding affect abstention in different ways, where larger or more capable models sometimes perform worse at timely abstention. Finally, we introduce CONVOLVE, a context engineering method for improving agentic abstention that distills full interaction trajectories into reusable stopping rules. On WebShop, CONVOLVE substantially improves timely abstention without updating model parameters, raising Llama-3.3-70B's timely recall rate from 26.7 to 57.4. Our dataset and code are available at https://lhannnn.github.io/agentic-abstention

1 Introduction

The paper frames Agentic Abstention as deciding when tool-using agents should stop rather than continue acting under uncertainty. Across more than 28,000 tasks, agents often abstain too late or not at all, while CONVOLVE improves timely abstention without parameter updates.

  • Motivation: Agentic Abstention addresses when tool-using agents should stop acting rather than continue interacting with an environment.Unlike single-turn answer-or-abstain decisions, agents can act, answer, or abstain as their context evolves.
  • Benchmark and evaluation: 28,000 instructions span web shopping, terminal interaction, and question answering, evaluated with 13 LLM-as-agent systems and 2 agent scaffolds.The benchmark combines adapted and newly constructed abstention tasks across these scenarios.
  • Findings: 26.7% timely abstention recall and 83.2% overall abstention recall were achieved by the strongest web baseline within the 10-turn budget.Many agents eventually recognize infeasibility but use unnecessary interactions first.
  • Findings: Less than 40% average timely recall was achieved by all tested models, while most systems achieved less than 50% average abstention recall.Timely abstention means stopping at the first turn when enough information is available.

2 Agentic Abstention

Agentic Abstention is defined as recognizing when a task is infeasible and stopping rather than answering incorrectly or taking unnecessary actions. The paper models this as a partially observable sequential decision process with answer, abstain, and act choices.

  • Definition: Agentic Abstention means recognizing task infeasibility and abstaining instead of answering incorrectly or taking unnecessary actions.The definition concerns stopping under the current information state.
  • Sequential formulation: The POMDP action space is A = {ANSWER, ABSTAIN, ACT}, combining terminal completion, terminal non-progression, and information-gathering actions.ACT includes external actions such as searches or environment interactions.
  • Sequential formulation: Partial observability makes the agent’s decision depend on interaction history, including the instruction, prior actions, and environment observations.The latent task state includes whether the task is resolvable with available context and tools.
  • Evaluation setting: Episodes terminate after ANSWER or ABSTAIN, or when the predefined step budget is reached.Instances are labeled abstain-warranted when they cannot be reliably solved under the interaction setting, even after appropriate actions.
  • Evaluation setting: Clarification requests count as ABSTAIN and terminate the episode rather than forming a separate primary action.ABSTAIN therefore includes stopping to request missing information.

3 Datasets

The benchmark covers web shopping, terminal execution, and interactive QA, combining solvable tasks with request-based and environment-based abstention variants. These variants make infeasibility explicit in the instruction or reveal it through interaction.

  • TerminalBench: TerminalBench 2.0 tasks combine containerized environments, instructions, verification tests, and reference solutions before instruction rewriting creates abstention variants.The adapted evaluation set contains 89 solvable and 188 unsolvable instances.
  • Benchmark scope: The benchmark evaluates web decision making, terminal task execution, and interactive QA while retaining solvable and abstain-warranted instances.The three scenarios use simulated shopping, command-line environments, and search-based question answering.
  • WebShop: 249 WebShop tasks use request-based abstention categories of Subjective Preference, Underspecified Intent, and False Premise or Contradiction.Instructions are manually reviewed after LLM-generated rewrites.
  • TerminalBench: Terminal request-based variants include 87 False Premise or Contradiction tasks and 80 Underspecified Intent tasks from 89 original tasks.A dual-agent correction pipeline revises invalid rewrites for up to three rounds before manual review.
  • Interactive QA: Interactive QA retains 16 AbstentionBench datasets, yielding 27,073 samples and converting each instance into answer, abstain, or search decisions.A fixed Wikipedia dump provides the retrieval corpus instead of live search APIs.

4 Experiment Setup

The experiments evaluate recent open-weight and API-accessed language models with Codex CLI and Terminus 2 across the benchmark scenarios. Abstention is measured by recall within a specified number of steps, with timely recall requiring abstention at the earliest warranted step.

  • Systems: The study evaluates 13 LLM-as-agent systems, including GPT-5.4-mini, Grok 4.1 Fast, Llama 3.3, Qwen 3, and other recent models.It also evaluates Codex CLI and Terminus 2 as agent scaffolds.
  • Systems: Proprietary models use first-party APIs, while open-weight models are accessed through Together AI and OpenRouter.
  • Metrics: AbsRec@K is the proportion of abstain-warranted instances where the agent abstains within K steps after abstention becomes warranted.Premature abstentions before the earliest warranted step are not counted as correct.
  • Metrics: Timely Recall equals AbsRec@1 for request-based abstention and AbsRec@2 for environment-based abstention.The distinction reflects whether infeasibility is evident initially or only after one interaction.
  • Metrics: AbsRec@10 is Overall Recall under the predefined maximum of Kmax=10 steps.The metric captures eventual correct abstention within the interaction budget.

5 Our Proposed Method for Improving Agentic Abstention: CONVOLVE

CONVOLVE improves agentic abstention by learning reusable stopping rules from multi-step interaction trajectories and adding them to the agent’s evolving context. It is instantiated on WebShop with fixed tool protocols and evaluated against a no-playbook baseline.

  • CONVOLVE framework: CONVOLVE analyzes full environment rollouts and distills abstention-relevant feedback into reusable stopping rules for future episodes.The framework updates an evolving playbook rather than learning only from final answers.
  • CONVOLVE framework: At episode k, the agent produces a trajectory containing the instruction, observations, and actions, then updates context using trajectory-derived feedback.The update is represented as c(k+1) = U(c(k), τ(k), y(k)).
  • Task setting & training configuration: CONVOLVE is trained on 20 WebShop abstention examples, validated on 8, and evaluated on 101 held-out examples from a 500-example abstention-only subset.The subset covers subjective preference, underspecified intent, false premise or contradiction, and missing target scenarios.
  • Environment rollouts: Rollouts keep the WebShop tool schema, execution protocol, task instructions, and tool-calling rules fixed while appending the learned playbook to the system prompt.Each rollout records observations, actions, termination reason, reward, invalid actions, abstention step, and timeliness.
  • Evaluation & baseline: The evaluation compares the same 70B model with CONVOLVE’s playbook injection against the same model without playbook injection under otherwise unchanged settings.The baseline uses the same examples, provider, temperature, token limit, interaction limit, timeout, and retry policy.

6 Results

Across web shopping, terminal interaction, and QA, agentic abstention remains difficult primarily because agents often recognize infeasibility too late. Abstention varies by category, reasoning effort, model scale, and context-engineering support.

  • Abstention performance: Agents often abstain only after unnecessary interaction, with timely recall lower than eventual recall across web, terminal, and QA settings.The benchmark evaluates whether agents abstain when infeasibility becomes evident and whether they do so at the earliest warranted step.
  • Abstention performance: Missing Target in Web, Underspecified Intent in Terminal, and False Premise and Underspecified Intent in QA are the most difficult abstention categories.Performance for the same model varies significantly across categories.
  • Factors impacting abstention: More reasoning can improve AbsRec@1 while reducing AbsRec@10, so additional reasoning does not uniformly improve abstention.For terminal tasks, medium reasoning can provide the best tradeoff, while high reasoning brings no further gains for GPT-5.4-mini.
  • Factors impacting abstention: Over-abstention increases with interaction on solvable tasks, especially in Web, while stronger reasoning generally mitigates false abstention.Qwen3-235B-Instruct reaches 34% over-abstention by turn 10, whereas medium- and high-reasoning GPT-5.4-mini variants remain at 0–2% in Terminal.
  • Factors impacting abstention: Scaling improves overall recall but not timely recall, indicating that larger models mainly become better at eventually abstaining rather than abstaining earlier.The scaling analysis varies Qwen model size and finds only slight changes in timely recall.
  • Improving abstention with CONVOLVE: CONVOLVE improves abstention with only 20 interaction trajectories, raising Llama-3.3-8B timely recall from 6.9 to 12.9 and overall recall from 92.1 to 94.1.Lessons learned by smaller models can also transfer to larger models and improve their performance.

7 Conclusion

The paper concludes that reliable agents need to judge when continued action is no longer useful, not only complete tasks. CONVOLVE improves timely abstention without updating model parameters.

  • CONVOLVE improves abstention on WebShop, with trained variants using 20 interaction trajectories.
  • Agents usually abstain too late or not at all, expending many wasteful actions when tasks become unresolvable only after environmental interaction.This occurs across web, terminal, and interactive QA settings.
  • CONVOLVE distills interaction trajectories into reusable stopping rules that can be provided as agent context without parameter updates.

Supplementary materials outline

The supplementary materials organize the paper’s limitations, broader implications, related work, datasets, and abstention scenarios across dedicated appendices.

  • Appendix A discusses study limitations, while Appendix B presents potential broader implications.
  • Appendix C covers related work on LLM abstention, agentic abstention, and context engineering.
  • Appendix D describes the datasets and their adaptations for question-answering, web, and terminal environments.
  • Appendix E defines the abstention scenarios considered in each evaluation setting.

A Limitations

The study’s benchmark and model coverage represent only portions of the broader agent landscape, and its constructed tasks may omit important forms of real-world infeasibility. Results should therefore be interpreted within the evaluated settings, systems, and public-benchmark context.

  • The benchmark covers web, terminal, and retrieval-augmented QA, but not richer interfaces, private tools, long-lived user state, or multi-step workflows.
  • Constructed abstention tasks may miss infeasibility from permission boundaries, stale resources, conflicting tool outputs, or prior user context.The environment-based tasks mainly focus on missing targets or prerequisites.
  • Findings are limited to the evaluated models and scaffolds rather than every reasoning setting or tool interface.Abstention behavior can depend on both the base model and the agent scaffold.
  • Public-benchmark exposure cannot be fully ruled out, despite evaluation across multiple scenarios and newly constructed variants.
  • Better abstention may reduce unnecessary tool use, but over-abstention can reduce usefulness and should be evaluated alongside task success.
  • The benchmark and method are intended for reliability research, not a complete deployment policy requiring no additional safeguards.Real-world systems may need clarification, escalation, human oversight, or domain-specific safeguards.

D.1.2 Dataset Adaptation

The benchmark adapts solvable web, terminal, and QA resources into abstention-warranted variants by rewriting instructions or modifying environments. It spans 16 datasets and 5 scenarios with 27,073 selected samples.

  • WebShop: 500 WebShop test instructions produce 500 additional abstention tasks through request-based and environment-based adaptations.Request-based cases are recognizable from the instruction; environment-based cases become recognizable as unresolvable after interaction.
  • WebShop: WebShop request-based instructions target Subjective Preference, Underspecified Intent, and False Premise or Contradiction, with author review.
  • WebShop: Environment-based WebShop tasks remove the ground-truth target item and rebuild the Lucene search index so the item cannot be retrieved.
  • Terminal-Bench 2.0: Terminal-Bench request-based variants use a Dual-Agent Correction pipeline with rewriting, validation, feedback, and up to three revision rounds.
  • Terminal-Bench 2.0: Terminal-Bench environment-based variants alter essential files, datasets, packages, configurations, or executable components across 21 manually selected tasks.
  • Question answering: The QA component uses 16 AbstentionBench datasets, with each sample containing a prompt, abstention label, and optional reference answer.
  • Question answering: QA evaluation provides search as a third action alongside answering and abstaining, while excluding datasets whose questions are often search-resolvable.
  • Question answering: The selected AbstentionBench subset contains 16 datasets and 5 scenarios, totaling 27,073 samples.

E Abstention Task Categories

The benchmark distinguishes request-based cases, where instructions are ambiguous or impossible, from environment-based cases, where interaction reveals missing targets or prerequisites. Abstention is evaluated by whether and how promptly agents stop after abstention becomes warranted.

  • Web environment: Web tasks include request-based contradictions, subjective preferences, and underspecified intent, plus environment-based Missing Target cases.Missing Target tasks initially appear solvable but become unresolvable when no valid item can be found.
  • Terminal environment: Terminal tasks include request-based contradictions and underspecified intent, plus Missing Prerequisite cases revealed through interaction.Missing prerequisites include absent files, dependencies, permissions, services, or external capabilities.
  • Question answering: Interactive QA covers Answer Unknown, False Premise, Subjective, Underspecified Context, and Underspecified Intent categories.The adapted categories are illustrative rather than exhaustive or mutually exclusive.
  • Metric formulation: Agents may answer, abstain, or act, and abstention is warranted only from the oracle step when sufficient evidence has emerged.Premature abstentions count as failures, preventing credit before the task supports stopping.
  • Metric formulation: AbsRec@K measures correct abstention within the first K interaction steps, while overall recall uses K=10 and timely recall uses K=1.Timely recall captures abstention at the earliest warranted step; when all warranted steps equal one, it is first-turn recall.
  • Metric formulation: SPL rewards successful abstention while penalizing delay, assigning zero to missed or premature abstentions and more credit to earlier successful decisions.For successful abstentions, the score compares the actual abstention step with the oracle earliest warranted step.

G.2 Results

Additional evaluations report that CONVOLVE improves abstention across question-answering and terminal settings, while qualitative examples expose failures from insufficient grounding and excluded false-premise cases.

  • Additional evaluations: CONVOLVE consistently improves abstention over base models on AbstentionBench and TerminalBench.On TerminalBench, Llama-3.3-70B gains in AbsRec@1, AbsRec@10, and SPL, with the strongest variant best across all three metrics.
  • Qualitative analysis: FalseQA is excluded from the main quantitative evaluation because correcting a false premise and answering can be a reasonable response rather than poor abstention.The paper presents representative FalseQA examples qualitatively instead.
  • Qualitative analysis: A QA example shows the agent answering “herd” after searches returned passages about American bison, local history, and European bison rather than the requested species.The example illustrates answering despite insufficiently grounded evidence.
  • Qualitative analysis: A terminal example shows the agent replacing unavailable tokenizer metadata with an assumed vocabulary size of n = 100256 instead of abstaining.The agent could not recover grounding because relevant project files and tokenizer libraries were unavailable.
  • Interactive protocol: The interactive protocol permits ANSWER, ABSTAIN, or SEARCH, with a bounded search budget and abstention required when evidence remains insufficient or unreliable.The protocol repeatedly exposes the original question, search budget, prior query, and retrieved passages before the next action.

J CONVOLVE Example Playbooks

CONVOLVE turns interaction trajectories into compact, reusable stopping rules that guide later agents without storing full trajectories. The learned playbook targets recurring signals of unfulfillable requests and unnecessary interaction.

  • CONVOLVE playbooks: CONVOLVE’s playbook combines general abstention principles with trajectory-derived guidelines learned from prior interactions.The guidance covers subjective, underspecified, and otherwise unfulfillable requests.
  • CONVOLVE playbooks: Rather than storing full trajectories, CONVOLVE appends compact stopping rules to later agent contexts while preserving normal tool use on resolvable tasks.The method is described as context engineering rather than parameter updating.
  • Agent scaffolds: With the same GPT-5.4-mini base model, Codex CLI achieves higher abstention recall than Terminus 2 across request-based and environment-based tasks.The comparison illustrates that scaffold choice affects abstention beyond the underlying model.
Loading 2606.28733v1…