Study finds LLM agents may need to share reasoning, not just do more of it
Researchers introduce information disclosure as a measurable skill for user-service agents and report that internal thinking alone does not reliably improve what agents tell users.
In plain words
Some AI assistants can reason privately before answering, but a new study says that silent reasoning does not guarantee users get the information they need. The authors measure how well agents disclose things like constraints and options, and find that prompting agents to share key details helps more than making them think longer.
What was announced
In a paper on arXiv, researchers study a communication-level failure mode in user-service LLM agents: internal reasoning can be counterproductive for interactive information disclosure. The authors treat disclosure as a measurable mechanism covering state corrections, available options, constraints, consequences, and proactively retrieved environment information, and test agents equipped with a "think" tool that lets them reason internally before responding or acting.
Across five models and three user-service scenarios, the paper reports that enforced thinking has unstable utility, and that more thinking does not reliably translate into richer disclosure. Using turn-level counterfactual intervention, the authors found that injecting golden decision-relevant information steered agents to correct action trajectories, suggesting missing disclosure causally affects downstream behaviour. Failed trajectories most often lacked state corrections and constraints, and a lightweight disclosure-aware prompt mitigated degradation. The authors conclude that the challenge is ensuring reasoning is externalised as timely disclosure, not simply making agents think more.
What it does not show
- The evidence comes from the paper's abstract only, so experimental setup details, scenario definitions and full results are not available for verification.
- The findings are the authors' own reported results across five models and three scenarios; no independent confirmation is implied.
For security and IT teams
- Scope
- The study covers user-service LLM agent scenarios where the agent holds information the user cannot see, such as task policies, tool results and environment states.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.