RSE-Bench
Benchmarking Skill Self-Evolution in Agents
RSE-Bench evaluates LLM agent skill self-evolution robustness under non-adversarial noise across four domains and six benchmarks, measuring retained performance gains.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info rse-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
8 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 95.4 | alfworld ace | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 95.4 | alfworld rethinkskill | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 95.4 | alfworld skillopt | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.2-chat | 84.2 | alfworld ace | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.2-chat | 84.2 | alfworld rethinkskill | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.2-chat | 84.2 | alfworld skillopt | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 43.3 | skillflow ace | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.2-chat | 32.7 | skillflow ace | Independent | medium | Oct 14, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
RSE-Bench spans spreadsheet manipulation, document understanding, interactive decision-making, and skill reuse. It injects controlled noise into instructions, evidence, records, and outcome pairings during the learning loop.
Scoring uses native rules on held-out tasks with frozen skills. Noisy learning retains only 25.5%–42.1% of clean evolution gains, with execution record noise causing significant losses.