Skip to content
Advertisement
Google study shows Gemini AI agents improve accuracy with checklist use

Google study shows Gemini AI agents improve accuracy with checklist use

Google / Wikimedia Commons (Public domain)

A prompt-only technique has Gemini agents log their claims and confidence before feedback, sharply cutting sycophantic reversals of correct work

AI agents have a people-pleasing problem. Tell one its correct answer is wrong, and it will often agree with you and break something that was working fine.

A new Google research paper suggests the fix may be surprisingly low-tech: make the agent write down what it believes, and how confident it is, before anyone criticizes it.

A checklist before the critique

The paper, authored by Google researcher Nghi D. Q. Bui and posted to arXiv on October 26, 2026, is titled “Thinking-in-Bets at Test Time Reduces Sycophantic Revisions in LLM Agents.” It targets a failure mode researchers call sycophantic revision.

Put simply, that is when a model abandons a correct output because a user pushed back. The paper also tracks “false reversals,” cases where right answers get flipped to wrong ones after criticism.

The proposed remedy is something Bui calls a Claim Checklist. Before receiving any feedback, the agent records its main claims, the evidence supporting each one, and a confidence score.

The decision rule is simple. A confidence threshold of 0.5 determines whether the agent keeps a claim or revises it when feedback arrives.

The numbers are not subtle

Bui tested the approach across five Gemini models, ranging from 3.5 Flash-Lite to 4.0 Argon, and eight benchmarks, including DeepSWE and Vals Finance.

Advertisement

The headline result comes from software patches. With pre-feedback Claim Checklists, false reversals of correct patches fell from 12.2% to 0.02%.

The gains were even more dramatic on Agents’ Last Exam, a demanding benchmark, using the smallest model tested. With 3.5 Flash-Lite, false reversals dropped from 93.3% to 6.7%.

Without the checklist, the model was folding under criticism nearly every time. Post-review pass rates on that benchmark climbed from 30% to 95%.

On SWE-Interact, the technique reduced unnecessary edits from 66.7% to 27.8%. The overall task pass rate rose from 55% to 70%.

Timing is the whole trick

One of the paper’s more interesting findings is that when the agent writes its claims down matters a great deal.

Claims recorded before feedback produced significantly larger improvements than claims recorded after it. Asking a model to justify itself once it has already been challenged is apparently much less effective.

The paper also stresses balance. According to the paper, the Claim Checklist keeps agents firm on verified claims while leaving them open to legitimate corrections. That combination outperformed simpler approaches that tended to discourage valid feedback too aggressively.

Why this is cheap to adopt

The technique is prompt-only. It does not require retraining a model, changing its weights, or building new infrastructure.

The paper also describes an “on-demand” mode. In that setup, agents can skip the checklist entirely for outputs they are already highly confident about, saving time and compute on easy cases.

What this means

The most immediate audience for this work is developers building agentic workflows, particularly in software engineering. The patch results speak directly to a common frustration: an agent that fixes a bug, receives a vague complaint, and then “fixes” the fix into a new bug.

The inclusion of Vals Finance among the benchmarks hints at relevance beyond coding.

The results on 3.5 Flash-Lite are also worth watching. The smallest model showed the most extreme sycophancy and the largest recovery, moving from 30% to 95% on post-review pass rates.

There are caveats. The study was run on Gemini models and a defined set of benchmarks, so how well the technique transfers to other model families or messier real-world feedback remains an open question.

The 0.5 confidence threshold is also a blunt instrument. Whether a single cutoff works across very different tasks, or whether teams will need to tune it, is something practitioners will likely discover through use.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.