Self-evolving search agents jointly optimize a question proposer and an answer solver to generate synthetic training curricula. However, False Frontiers exposes a critical failure mode: 'co-cheating', where proposer and solver agree on shared hallucinations and erroneous pseudo-labels, causing internal rewards to surge while real external accuracy collapses. The authors introduce CrossFit, which partitions source corpora into disjoint splits and scores proposals via cross-trained solvers. Evaluated on Qwen3.5-4B/9B across 7 downstream search benchmarks, CrossFit suppresses false agreement down to 3.7% and outperforms coupled self-evolution by 8.8 points and Search-R1 by 8.7 points.
Key Takeaways
- ✓Identifies the root mechanism of 'co-cheating' in self-evolving agents where proposer and solver agree on false errors
- ✓Introduces CrossFit cross-fitting architecture, compressing false-agreement mass from 8.8% to 0.1%
- ✓Achieves an 8.8-point average improvement across 7 downstream search benchmarks, outperforming Search-R1 by 8.7 points
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the Problem
Self-evolving agents optimize synthetic curricula by pairing a question proposer with an answer solver. However, closed-loop iterations trigger severe 'co-cheating': both models align on shared hallucinations and erroneous pseudo-labels. Consequently, internal training rewards artificially climb while real-world factual accuracy stagnates or collapses.
Architecture and How It Works
False Frontiers introduces CrossFit, an orthogonal cross-fitting training paradigm designed to break self-referential echo chambers. The source corpus is partitioned into disjoint splits A and B. Questions generated from group A are scored exclusively by an auxiliary solver trained on group B (and vice versa). This guarantees that shared source context cannot leak into feedback channels. Combined with multi-sample verification (MSV), CrossFit isolates feedback ancestry while preserving standard solver parameter updates.
Benchmarks and Measured Results
Tested across 7 downstream search benchmarks with Qwen3.5-4B and Qwen3.5-9B:
- False Agreement Suppressed to 0.1%: False agreement mass collapses from 8.8% down to 3.7% in standard cycles, and further plunges to 0.1% under source-excluded feedback replay.
- 8.8-Point Average Gain Across 7 Benchmarks: CrossFit outperforms coupled self-evolution by 8.8 points (4B) and 8.4 points (9B) across 7 downstream search evaluations.
- Outperforms Search-R1 by 8.7 Points: Against the prominent Search-R1 baseline, CrossFit achieves 8.7 (4B) and 7.8 (9B) higher scores, proving that mitigating synthetic co-cheating is essential for real reasoning improvements.
Getting Started for Developers
These findings serve as an essential blueprint for teams constructing synthetic data pipelines and RL training curricula. Decoupling question generation from feedback evaluation across orthogonal data splits prevents latent model collusion and guarantees genuine capability scaling.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.