Google study shows Gemini AI agents improve accuracy with checklist use

2 weeks ago 33

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

Alphabet Inc. (GOOGL) price

Full chart and stats

Loading chart…

Market cap $3.5T (shares as of 2026-09-30)

P/E (trailing) 14.8 (EPS to 2026-09-30)

Annual revenue $402.8B USD (FY ended 2025-12-31) 10-K
Annual revenue growth +15.1% (FY ended 2025-12-31 vs FY ended 2024-12-31) 10-K
Annual net income 132,170,000,000 USD (FY ended 2025-12-31, attributable to the company) 10-K

Annual operating margin 32.0% (FY ended 2025-12-31)

Forward P/E Unavailable: no dated forecast source

Gloria daily close for session Oct 29, 2026. Market cap and trailing P/E use this close with the SEC share total and trailing diluted earnings, each dated next to its figure. Annual revenue, revenue growth, net income and margin show their own fiscal period and link the 10-K. Current context may postdate this article.

Alphabet Inc. and its peers

Token vs underlying stock

Current comparison for GOOGL. Historical observations may postdate this article.

GOOGLx token unit vs GOOGL share · USD

$284.45Oct 29 UTC

-20.6% over displayed period

● GOOGLx token┄ GOOGL stock: $283.82

Aug 6, 2026: token $358.14 at 20:00 UTC; stock close $357.75 Aug 7, 2026: token $355.20 at 20:00 UTC; stock close $354.30 Aug 10, 2026: token $357.76 at 20:00 UTC; stock close $357.52 Aug 11, 2026: token $344.26 at 20:00 UTC; stock close $343.80 Aug 12, 2026: token $344.44 at 20:00 UTC; stock close $343.54 Aug 13, 2026: token $347.06 at 20:00 UTC; stock close $346.36 Aug 14, 2026: token $346.47 at 20:00 UTC; stock close $345.90 Aug 17, 2026: token $344.71 at 20:00 UTC; stock close $344.00 Aug 18, 2026: token $344.71 at 20:00 UTC; stock close $344.20 Aug 19, 2026: token $345.42 at 20:00 UTC; stock close $344.72 Aug 20, 2026: token $341.21 at 20:00 UTC; stock close $340.67 Aug 21, 2026: token $345.60 at 20:00 UTC; stock close $344.82 Aug 24, 2026: token $348.94 at 20:00 UTC; stock close $348.06 Aug 25, 2026: token $347.59 at 20:00 UTC; stock close $346.96 Aug 26, 2026: token $342.69 at 20:00 UTC; stock close $342.00 Aug 27, 2026: token $341.18 at 20:00 UTC; stock close $340.65 Aug 28, 2026: token $347.10 at 20:00 UTC; stock close $346.59 Aug 31, 2026: token $339.89 at 20:00 UTC; stock close $339.35 Sep 1, 2026: token $335.42 at 20:00 UTC; stock close $335.02 Sep 2, 2026: token $337.78 at 20:00 UTC; stock close $337.12 Sep 3, 2026: token $342.96 at 20:00 UTC; stock close $342.48 Sep 4, 2026: token $339.38 at 20:00 UTC; stock close $338.46 Sep 8, 2026: token $338.99 at 20:00 UTC; stock close $338.36 Sep 9, 2026: token $331.31 at 20:00 UTC; stock close $330.65 Sep 10, 2026: token $333.00 at 20:00 UTC; stock close $332.60 Sep 11, 2026: token $339.44 at 20:00 UTC; stock close $338.50 Sep 14, 2026: token $350.02 at 20:00 UTC; stock close $349.39 Sep 15, 2026: token $345.54 at 20:00 UTC; stock close $344.98 Sep 16, 2026: token $343.72 at 20:00 UTC; stock close $342.87 Sep 17, 2026: token $348.19 at 20:00 UTC; stock close $347.33 Sep 18, 2026: token $349.96 at 20:00 UTC; stock close $349.54 Sep 21, 2026: token $355.68 at 20:00 UTC; stock close $354.97 Sep 22, 2026: token $351.84 at 20:00 UTC; stock close $351.16 Sep 23, 2026: token $338.83 at 20:00 UTC; stock close $337.83 Sep 24, 2026: token $343.15 at 20:00 UTC; stock close $342.36 Sep 25, 2026: token $344.56 at 20:00 UTC; stock close $343.92 Sep 28, 2026: token $343.44 at 20:00 UTC; stock close $342.75 Sep 29, 2026: token $341.65 at 20:00 UTC; stock close $340.92 Sep 30, 2026: token $344.99 at 20:00 UTC; stock close $344.08 Oct 1, 2026: token $339.14 at 20:00 UTC; stock close $338.24 Oct 2, 2026: token $344.17 at 20:00 UTC; stock close $343.50 Oct 5, 2026: token $347.05 at 20:00 UTC; stock close $346.47 Oct 6, 2026: token $348.53 at 20:00 UTC; stock close $347.68 Oct 7, 2026: token $351.27 at 20:00 UTC; stock close $350.50 Oct 8, 2026: token $348.51 at 20:00 UTC; stock close $348.29 Oct 9, 2026: token $352.71 at 20:00 UTC; stock close $351.66 Oct 12, 2026: token $353.57 at 20:00 UTC; stock close $352.72 Oct 13, 2026: token $354.45 at 20:00 UTC; stock close $353.78 Oct 14, 2026: token $347.97 at 20:00 UTC; stock close $347.64 Oct 15, 2026: token $350.19 at 20:00 UTC; stock close $349.22 Oct 16, 2026: token $341.50 at 20:00 UTC; stock close $340.59 Oct 19, 2026: token $345.25 at 20:00 UTC; stock close $344.18 Oct 20, 2026: token $340.61 at 20:00 UTC; stock close $340.28 Oct 21, 2026: token $335.85 at 20:00 UTC; stock close $334.76 Oct 22, 2026: token $324.33 at 20:00 UTC; stock close $324.22 Oct 23, 2026: token $321.30 at 20:00 UTC; stock close $320.21 Oct 26, 2026: token $317.69 at 20:00 UTC; stock close $316.80 Oct 27, 2026: token $321.43 at 20:00 UTC; stock close $320.58 Oct 28, 2026: token $325.97 at 20:00 UTC; stock close $324.89 Oct 29, 2026: token $284.45 at 20:00 UTC; stock close $283.82 Aug 6$283.82 to $358.14Oct 29

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.

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.

Read Entire Article