AI

More Rules, Longer Lists? Benchmarking Self-Evolving Agents on a Hidden-Rule Curriculum

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/11/2026, 05:03:24 UTC.

Briefings

A new benchmark named CLR-Bench evaluates self-evolving agents on a curriculum where hidden rules are added incrementally. Six recent methods were tested, with GPT-6-Sol and ACE showing significant gains over frozen agents. The study highlights that agents learn from feedback alone but struggle with unseen members, while a fixed lesson significantly improves performance.

Key points

• CLR-Bench introduces a curriculum of latent rules where each chain adds one hidden rule per level on top of earlier ones.
• The best self-evolving methods reached 78% accuracy and solved 59-60% of single cells, significantly outperforming frozen agents.

Context

This benchmark tests whether self-evolving agents can learn a rule nobody told them, using a curriculum of latent rules. It measures how well agents learn from feedback versus raw history, with a fixed lesson providing a baseline improvement.

Limitations

The study evaluates six methods on 20 chains across 10 domains. Uncertainty exists regarding the exact number of unseen members solved and the precise impact of the fixed lesson on unseen members.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/11/2026, 06:14:09 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeick7fjrlype7kr2bopd7qlus2hwm4d6apj3g7ncorfkr7g2lvo3cu