HMED Posts Highest Mean Results Across Six AI Benchmark Combinations

Five researchers tested a method for improving how AI agents discover task skills, and it posted the highest mean result across all six model-benchmark combinations.

Their Hindsight Meta-Experience Distillation (HMED) approach compares an agent’s old and revised skill-discovery strategies after returning to the same starting state. The authors submitted a preprint describing the method on Oct. 6, 2026.

HMED focuses on “Meta-Skills,” the strategies an agent uses to find and improve task-specific skills. When one of these strategies changes, the method restores the state that prompted the change and runs both versions from there. It saves the comparison as information that can guide later updates, including when an agent does not keep the revised strategy.

Across ALFWorld, ScienceWorld and BiomniBench-DA, the team evaluated HMED using Qwen3-32B and DeepSeek-v4-Flash. The first two benchmarks measure task success rates; BiomniBench-DA uses a judge score on a 100-point scale. Main results were averaged across three independent runs and compared with MetaSkill-Evolve, a baseline that also develops Meta-Skills.

With Qwen3-32B, HMED scored 54.11% on ALFWorld, against 51.49% for MetaSkill-Evolve. On ScienceWorld, the scores were 22.56% and 20.84%; on BiomniBench-DA, they were 20.78 and 17.80 out of 100. With DeepSeek-v4-Flash, HMED scored 80.60% versus 76.49% on ALFWorld, 32.47% versus 28.70% on ScienceWorld, and 43.19 versus 36.80 on BiomniBench-DA.

A separate ALFWorld comparison examined whether using the strategy comparison in later updates added value. Full HMED scored 54.11%, while a version that made comparisons but did not pass their consequences into later updates scored 49.25% — a gap of 4.86 percentage points.

The benchmark results do not establish how HMED would perform in deployed products or whether it would reduce the need for human supervision.