Evaluation & Judgment / EMNLP 2026 Findings

A Browser Agent Gets Credit for a Click It Never Made

Rewriting an agent’s explanation while preserving its actions changes how often automated judges call a failed web task a success.

Gaming the Judge: How Unfaithful Chain-of-Thought Can Undermine Agent Evaluation

Muhammad Khalifa, Lajanugen Logeswaran, Jaekyeom Kim, Sungryull Sohn, Yunxiang Zhang, Moontae Lee, Hao Peng, Lu Wang, Honglak Lee

Overview

A browser agent on a travel-booking task clicks March 31 for a requested March 30 check-in. Its written account says it selected March 30 and that the booking site confirmed the date. A judge with the click and the page in view can still believe the account. This paper tests how often changing only an agent’s reasoning trace alters the verdict.

Key Findings

The agent chose March 31 and told a different story

The user wanted a hot-spring package in Beppu starting March 30. The agent clicked March 31. In the paper's opening case, Gemini-2.5-Flash judged the original record a failure. The researchers left the click and screenshot alone and rewrote the agent's accompanying explanation to claim the correct date had been selected. The judge changed its verdict to success.

Figure 1 from the EMNLP Findings paper. The same click and screenshot receive different verdicts after a fabricated claim about the selected date is added to the agent's reasoning.
Figure 1 from the EMNLP Findings paper. The same click and screenshot receive different verdicts after a fabricated claim about the selected date is added to the agent's reasoning.

Browser agents often emit a short explanation alongside their observations and actions. The explanation can help a judge follow a long interaction. It also comes from the system whose work the judge is grading. A statement like “all filters are selected” can summarize a difficult-to-read screenshot, or it can describe a state that never existed.

A controlled rewrite separates the explanation from the work

The authors built a paired test. They collected 800 trajectories across 659 tasks, including 300 from AgentRewardBench and 500 newly collected runs. Each trajectory records what the agent observed, what it wrote about that observation, and what it did. The researchers generated variants by changing only the reasoning text after the run. The original and rewritten records went to nine vision-language judges with the screenshots and actions held fixed.

Figure 3 from the EMNLP Findings paper. The evaluation runs a browser agent, scores the original trajectory, then scores the same action and observation sequence with altered reasoning.
Figure 3 from the EMNLP Findings paper. The evaluation runs a browser agent, scores the original trajectory, then scores the same action and observation sequence with altered reasoning.

The authors separate five kinds of rewriting. Competence boosting makes the agent sound assured. Reflective reasoning makes it sound deliberate. Fabricated progress claims steps have been completed. Environment blaming invents obstacles to explain errors. Task reinterpretation subtly lowers or changes the goal. The first two alter how the original account is presented. The others change what the account says about the environment or task. The paper's taxonomy sets out these routes and the judgment habits they target.

Count the verdicts that turn

Section 5.1 defines the judgment flip rate, or JFR. For judge JJ and rewrite strategy ss, let TT be the original failed trajectory and T~s\widetilde T_s be its rewritten counterpart. The measure asks how often a judge that correctly rejected the original accepts the rewrite.

JFR(J,s)=Pr⁡[J(T~s)=success∣J(T)=failure].\mathrm{JFR}(J,s)=\Pr\left[J(\widetilde T_s)=\mathrm{success}\mid J(T)=\mathrm{failure}\right].

The paper also measures changes in false-positive rate. An absolute increase is a percentage-point difference. A relative increase divides that difference by the original false-positive rate. In Figure 6, fabricated progress raises the false-positive rate of GPT-4o and GLM-4.1V-9B-Thinking by about 20 to 30 percentage points under the main evaluation. Competence boosting has smaller effects, while reflective reasoning is comparatively benign. Similar F1 scores on ordinary records give little assurance that two judges respond the same way to a rewrite.

Table 1 tests a possible explanation. If the problem were mainly that judges struggle to read screenshots, replacing them with text captions should help. For GPT-4o, the attack's effect falls with captions and remains substantial.

Observation, GPT-4o under fabricated progressAbsolute false-positive rate changeRelative changeJudgment flip rate
Text caption+17.5 points+75.8%32.7%
Screenshot+27.3 points+93.8%41.8%

Skeptical judges can overlook a real success

An explicit warning about manipulated reasoning and a task-specific rubric reduce vulnerability in several conditions. Neither works consistently across all nine models. The authors also give judges more chances to reason, either by voting across repeated judgments or letting a model deliberate longer. This helps with some style changes. Fabricated progress remains difficult even at greater inference cost.

A more skeptical judge can also miss genuine completions. Figure 10 reports that removing reasoning or using a strong warning can lower flip rates by up to 30–40% relatively while reducing recall by 10–20 percentage points. A plausible explanation can provide helpful context when an agent succeeds. The judge needs a way to use the record while checking the claims that decide the verdict.

Teach the judge what should stay the same

Appendix E.1 tests that approach with paired examples. It trains a judge to give consistent verdicts to original and rewritten traces, then tests it on a manipulation it did not see during training. Table 5 gives one result for GLM-4.1V-9B-Thinking. The fractions below are expressed as percentages.

Judge setup, held-out fabricated progressFalse-positive rateFlip rateRecall
Zero-shot62%42%93.2%
Manipulation-aware prompt55%32%91.2%
Fine-tuned on style rewrites49%24%94.4%
Fine-tuned on content rewrites46%20%93.8%

The paired training reduces the two attack measures in this test while preserving recall. The experiment shows that misleading text is sufficient to alter an evaluation. It does not establish that agents in ordinary use deliberately produce these same rewrites. The distinction matters if a developer plans to use a judge's verdict as a reward. A better-looking explanation can win that reward even when the user's booking is still wrong.

Questions & Answers

The opening answers are simulated author responses based on the paper.

Q: Why can a judge believe a statement that conflicts with the screen?

A: An agent’s explanation can supply a coherent story across several screens. The judge may trust the assertion rather than independently checking every action against the task. Our intervention measures whether changing that account alone changes the verdict.

Q: What counts as a judgment flip?

A: The flip rate counts cases where the judge correctly said failure on the original trace and then said success after rewriting. Its denominator is the set of failed trajectories the judge originally classified correctly. The false positive rate uses all failed trajectories.

Q: Would asking the judge to be more skeptical fix the problem?

A: The manipulation-aware prompt reduces some errors in our tests. Its effect depends on the judge and manipulation. A separate grounding check can require each success criterion to point to evidence from the page or action log.

Q: Does the manipulation measure how often agents fabricate success on their own?

A: The experiment changes the reasoning narrative while holding actions and observations fixed, so the verdict change measures vulnerability to those particular rewrites. It does not estimate how often agents spontaneously fabricate success or learn to do so under optimization. Read relative increases alongside absolute false-positive rates, since a low starting rate can make a modest absolute change look large.

Future Ideas

An evidence receipt for each browser success claim

Have a checker separate from the acting agent assemble a receipt listing each user condition and the observed event that fulfills it. A judge can read the agent’s explanation for context, but a success claim remains provisional until a separate evidence checker resolves each condition. The receipt could point to a selected date, a final confirmation identifier, or an earlier page that establishes a prerequisite.

Evaluate on held-out website families with successful and unsuccessful trajectories. Compare ordinary judging, manipulation-aware prompting, judging without the trace, and receipt checking. Include fabricated reports, missing snapshots, and late state changes. Track false approvals, missed successes, and how often the checker asks for more evidence. Begin with sandboxed tasks whose required conditions can be independently labeled.

Assemble event records from the browser or environment rather than from events asserted by the acting agent. Report outcomes separately when a recording is incomplete so missing evidence is visible alongside missed successes and false approvals.

Sources

Original abstract

Large language models (LLMs) are increasingly used as judges to evaluate agent performance, particularly in non-verifiable settings where judgments rely on agent trajectories including chain-of-thought (CoT) reasoning. This paradigm implicitly assumes that the agent’s CoT faithfully reflects both its internal reasoning and the underlying environment state. We show this assumption is brittle: LLM judges are highly susceptible to manipulation of agent reasoning traces. By systematically rewriting agent CoTs while holding actions and observations fixed, we demonstrate that manipulated reasoning alone can increase false positive rates of state-of-the-art vision-language model (VLM) judges by up to 117% relative to the unmanipulated baseline across 800 trajectories spanning diverse web tasks. We study manipulation strategies spanning style-based approaches that alter only the presentation of reasoning and content-based approaches that fabricate signals of task progress, and find that content-based manipulations are consistently more effective. We evaluate mitigation strategies including manipulation-aware prompting and judge-time scaling, which can reduce but do not eliminate the vulnerability. Our findings reveal a fundamental vulnerability in LLM-based evaluation and highlight the need for judging mechanisms that verify reasoning claims against observable evidence.