What question did this study set out to answer?

This research aims to investigate the stability of self-evaluation in self-improving language models over multiple training iterations.

May 1, 2026Open Access

Verification Collapse in Iterative Self-Improving Language Models

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial investigates self-evaluation accuracy in self-improving language models, revealing verification collapse.

This research aims to investigate the stability of self-evaluation in self-improving language models over multiple training iterations.
Analyzed self-evaluation accuracy across 20 iterations using the MATH benchmark with three model families.
Introduced adversarial hard-negative injection to mitigate the verification gap.
Compared a ground-truth-based control to isolate the effects of self-scoring.
Verification gap increased across all models with Qwen2.5-7B showing Δ_1 = 0.173 and Δ_20 = 0.320.
Adversarial hard-negative injection reduced the verification gap by 47% on Qwen2.5-7B.
Control with ground-truth filtering eliminated collapse entirely, confirming self-scoring as the causal driver.

Leonardo Giuseppe Gianola (2026) studied this question.

View Full Paper

Ask AI

Bookmark

May 1, 2026Open Access

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial investigates self-evaluation accuracy in self-improving language models, revealing verification collapse.

This research aims to investigate the stability of self-evaluation in self-improving language models over multiple training iterations.
Analyzed self-evaluation accuracy across 20 iterations using the MATH benchmark with three model families.
Introduced adversarial hard-negative injection to mitigate the verification gap.
Compared a ground-truth-based control to isolate the effects of self-scoring.
Verification gap increased across all models with Qwen2.5-7B showing Δ_1 = 0.173 and Δ_20 = 0.320.
Adversarial hard-negative injection reduced the verification gap by 47% on Qwen2.5-7B.
Control with ground-truth filtering eliminated collapse entirely, confirming self-scoring as the causal driver.

Leonardo Giuseppe Gianola (2026) studied this question.

View Full Paper

Ask AI

Bookmark