Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near Claude Opus 4.5

Meta AI and the University of Illinois Urbana–Champaign say their EvoHarness-RL framework let Qwen3-8B reach 96.9% average success on ALFWorld, edging Claude Opus 4.5’s reported 96.4% score. The striking comparison is bounded: it comes from a text-based benchmark of multi-step tasks, while the framework’s larger pitch is for agents working across long, changing workflows.

The runtime layer becomes part of the training target

An agent’s harness is the runtime layer that supplies execution feedback, state tracking and control flow around the model. EvoHarness-RL treats management of that layer as a learned behavior rather than a fixed set of developer-written instructions. The approach asks the model to decide when it should consult, change or consolidate the information outside its context window.

Its unified workspace separates external state into Belief, meaning the agent’s current view of its environment; Progress, meaning completed and pending subgoals; and Experience, meaning knowledge carried across tasks. The model operates those areas through four meta-actions: track, commit, recall and note.

Tool use has a budget

The training recipe has two stages. Supervised harness fine-tuning teaches the base model to turn messy interaction logs into structured external state. Cost-aware reinforcement learning then teaches it to weigh the benefit of retrieving or updating that state against the added token and compute cost.

96.9%EvoHarness-RL-trained Qwen3-8B

The trained Qwen3-8B scored 96.9% average success on ALFWorld, 49.0 percentage points above its baseline ReAct counterpart.

96.4%Claude Opus 4.5

Claude Opus 4.5 scored 96.4% out of the box in the reported ALFWorld comparison.

A benchmark gain, with a broader but unproven deployment case

The Qwen3-8B result also exceeded the reported scores for SkillRL, at 89.9%, and SkillOS, at 80.2%. Researchers further found that adding the BPE harness at prompt time improved GPT-4.1 by 22.1 percentage points and GPT-5 by 25.7 points. Those figures suggest the workspace can help models beyond the one trained with EvoHarness-RL, but they remain results from the same reported evaluation.

The researchers describe two behaviors that explain the intended efficiency gain. During harness annealing, an agent reduces external-state queries as routine patterns are absorbed during training. During harness evolution, it increases use of Belief and Experience when it encounters a novel environment or a difficult situation.

Where the framework is meant to earn its keep

For short, stable tasks, the researchers say ReAct or standard retrieval-augmented generation may already be sufficient. They position BPE as more useful when an agent operates for hours, days or weeks and needs to keep an updated, compressed record of its environment, unfinished work and prior failures.

The implementation is designed to sit above domain-specific tools through an environment adapter, rather than require a wholesale replacement of an existing agent framework. The unresolved test is operational: whether the benchmark’s learned balance between memory access and action holds up when enterprise state is incomplete, changing and costly to consolidate.