When Models Agree, That Is Not Evidence They Are Right ── Co-cheating in self-evolving search agents, and CrossFit's split-source grading

1. The trap in letting a model write and answer its own questions

Training an agent that answers questions with the help of search requires a large supply of question-answer pairs. Writing them by hand is expensive, so researchers have turned to letting models generate the questions themselves, having another role solve them, and training both on the outcome. The self-evolving search agents studied in this paper jointly optimize a proposer, which writes questions from source documents, and a solver, which searches and answers them. In this way the system builds its own training curriculum.^[1]

Because no external answer key exists, "correct" inside this loop means that the solver's answer matches the answer the proposer had in mind. The authors argue that this creates a specific failure mode, which they call co-cheating: the proposer and solver increasingly agree on the same errors, so the internal reward rises while correctness, as seen from outside, does not. Internal metrics keep reporting progress while real performance stalls or falls, and that kind of gap is among the hardest to catch in practice. The primary source is a preprint posted on arXiv and has not been peer reviewed.

2. The core idea: cut the grader off from the question's source

The authors first test the most direct fix, which they call multi-sample verification (MSV). Proposals are verified before training: the same model is queried three times with the source and three times without it, and the results decide whether a task is admitted and whether an unreliable pseudo-label should be replaced. According to the abstract, MSV only partly reduces false agreement, leaves substantial co-cheating in place, and costs six extra labeling generations for every candidate.

Those limits motivate the main method, CrossFit. The proposer's source documents are split into two groups, A and B. Questions generated from A are scored by an auxiliary solver trained only on B, and questions from B are scored by one trained only on A. The proposer's reward is set by this cross-fitted agreement.

The aim is that a pseudo-label tied to one source cannot be reproduced by the solver giving feedback. If the proposer misreads a document and builds a wrong answer key from it, a grader that never saw that document is unlikely to arrive at the same mistake. The authors state that the original solver's update rule is unchanged; only the origin of the proposer's reward differs.

3. What the abstract reports

Using a post-hoc audit against source evidence held outside the training loop, the authors report that co-cheating became more severe over successive rounds of self-evolution. Pseudo-label correctness stagnated or declined even as the in-loop training signal kept improving.

The mitigations are compared by "false-agreement mass". Rerunning the loop with Qwen3.5-4B and Qwen3.5-9B, MSV lowered it from 6.1% to 5.7% and from 8.8% to 7.2%. CrossFit lowered it to 3.0% and 3.7%. Replaying the identical proposals while switching only the feedback to a source-excluded solver lowered false agreement further, to 0.4% and 0.1%. The authors present this replay as a way to separate the effect of where the feedback came from from the effect of a changed curriculum.

On downstream performance, averaged across seven search benchmarks, CrossFit is reported to beat standard coupled self-evolution by 8.8 and 8.4 points at 4B and 9B, and to beat the Search-R1 baseline^[3] by 8.7 and 7.8 points. These are results under the paper's experimental conditions and do not show that the same margins will appear elsewhere.

4. A concrete case: generating a Q&A set from internal documents

Consider a team that wants to train a search agent by automatically generating questions from internal procedures and product documentation. Suppose one procedure contains a sentence that can be read in two ways.

Both setups use the same yardstick: did the answers agree? The only difference is what the judge has already read, in other words, how independent the judgment is. Watching only the internal agreement rate, the left-hand case looks like steady improvement.

5. What is not new, and what the abstract leaves open

Splitting data, fitting on one part, and evaluating on another is not a new idea. In statistics it is known as cross-fitting, a standard way to avoid the bias that comes from using the same data both to estimate and to evaluate.^[4] The paper's contribution is better described as bringing that idea into the closed proposer-solver loop, and naming and measuring co-cheating as a distinct failure mode.

Several limits remain. First, CrossFit does not bring false agreement to zero. The fact that the source-excluded replay goes lower still suggests that some same-source agreement survives CrossFit's partition.

Second, if the same fact or the same error appears in both group A and group B, splitting the documents does not make the judgment independent. This is plausible for corporate document sets with many versions and copied passages, but the abstract does not address the condition.

Third, the abstract does not say how correctness was decided in the post-hoc audit, whether by human reviewers or by another model, and the audit method directly shapes the false-agreement figures. Fourth, the cost of training two auxiliary solvers is not stated in the way MSV's cost is. Fifth, only two sizes from one model family were tested, and the seven benchmarks are neither named nor broken out individually in the abstract.

6. Why this paper is getting attention, and the state of the cluster

In this collection cycle the paper was selected on its own. Its cluster size is 1; no paper from another group addressing the same question was found within the collection window. The selection rests on reader votes on the paper-sharing site, where it has 159 upvotes and 1 comment.^[2] By the design's own definition this is a single paper that caught attention, not a topic rising as a cluster. Upvotes indicate researcher interest, not that the claims are correct.

There is still a reason to cover it. Work on teaching models to use search through reinforcement learning has continued since Search-R1, and it is moving toward letting models produce their own training material to reduce reliance on human answer keys. The further that goes, the harder it becomes to tell "the internal metric is rising" apart from "the model is actually getting things right". This paper tries to show, round by round and with an external audit, how that distinction breaks down. That makes it worth reading as a diagnostic record as much as a method proposal.

7. Where this connects to pharmaceutical and regulatory work

Quality assurance rests on separating the person who does the work from the person who checks it. Two people who worked from the same assumptions and the same documents will not catch a shared misunderstanding by checking each other. CrossFit tries to build that independence of checking into the training loop itself.

When generative AI is used for product inquiry handling or for searching internal policies, it is tempting to have the same model generate the evaluation questions, simply because it saves effort. What this paper shows is that such a setup can produce evaluation scores that rise while correctness does not. Before an internal agreement rate is recorded as evidence of performance, it is worth confirming which documents the evaluation was built from and who, or which model, did the judging.

In the end, the reference point for correctness is the primary documentation, such as approved labeling and authorized materials, together with the judgment of the people accountable for reading it. The paper's central concern, that agreement between models should not stand in for correctness, applies directly to validation design for AI in regulated work, even allowing for the fact that this is research that has not yet passed peer review.