Meta AI's 8B Model Ties Claude Opus 4.5 on Agent Benchmark With EvoHarness-RL

Researchers at Meta AI and the University of Illinois Urbana-Champaign (UIUC) have unveiled EvoHarness-RL, a reinforcement-learning method for training AI agents. According to a report published by AI Times on August 29, the team applied the method to Qwen3-8B, an open-source model with just 8 billion parameters, and reached a 96.9% success rate on ALFWorld, a benchmark of sequential, multi-step household tasks. That result edges out the 96.4% Anthropic's flagship model, Claude Opus 4.5, scores with the plain ReAct approach, and marks a 49.0 percentage-point jump over Qwen3-8B's own ReAct baseline of 47.9%. VentureBeat covered the same research separately on August 28, reporting that OpenAI's GPT-4.1 and GPT-5 also improved by 22.1 and 25.7 percentage points respectively when the method was applied at inference time only, with no retraining.

BPE: splitting an agent's state into belief, progress and experience

At the center of EvoHarness-RL is a state abstraction the researchers call BPE — Belief, Progress, Experience. Belief captures what the agent currently knows about its environment; Progress tracks completed subgoals and remaining work; Experience stores skills, past failures and retrieval priorities gathered from earlier tasks. To manage these three states, the agent relies on just four meta-actions: track (check the current state), commit (log progress), recall (retrieve past experience) and note (record new information). The researchers said that where earlier external-memory and tool-use systems followed rules fixed in advance by humans, BPE trains the agent to decide for itself when to record something and when to pull up past experience.

Two training stages: supervised fine-tuning, then reinforcement learning

Training happens in two stages. In the first, supervised fine-tuning (SFT) stage, Qwen3-8B first learns what BPE means and how to use the four meta-actions. In the second, reinforcement-learning stage, the team applies what it calls cost-aware Group Relative Policy Optimization (GRPO), training the model to judge for itself whether calling the harness actually helps it complete a task. Per VentureBeat, co-author Xuying Ning said that the optimal harness often changes with the model, and that append-only memory assumes more context is always helpful, which is not necessarily true. When the researchers removed the Experience component in an ablation test, average success fell to 48.6% — the steepest drop of the three components.

• Plain ReAct baseline: 47.9%
• BPE applied at inference time only (no retraining): 56.4%
• With supervised fine-tuning (SFT): 68.6%
• With SFT plus reinforcement learning (EvoHarness-RL, final): 96.9%
• Claude Opus 4.5 — 96.4% with plain ReAct, 98.5% with BPE added (+2.1 points)
• GPT-4.1 — 47.9% to 70.0% with inference-time BPE (+22.1 points)
• GPT-5 — 60.7% to 85.0% with inference-time BPE (+25.7 points)
• Comparison systems — SkillRL 89.9%, SkillOS 80.2%

It held up in unfamiliar settings — and revealed 'harness annealing'

The team also tested the model on new ALFWorld tasks it had never seen during training. The base Qwen3-8B managed only 50.0%; adding BPE at inference time alone raised that to 77.6%; and the fully trained EvoHarness-RL model reached 86.6%. During training, the researchers observed a pattern they call harness annealing: early on, the agent called the external harness at nearly every step, but as training progressed, repeated task patterns became internalized, harness calls dropped sharply, and usage eventually settled at roughly once per episode. The Experience store evolved the same way — quickly accumulating a wide range of skills and mistakes early on, then later consolidating redundant entries and discarding low-value ones, so that it kept only essential information rather than simply growing without limit. In one example the researchers cited, an agent searching for a kettle found that its stored memory of the counter's location conflicted with the actual environment, and corrected itself by writing a new note flagging the old information as wrong.

From software development to financial compliance

The researchers argued the framework extends beyond game-like environments into real work. In software development, Belief would track a codebase's current state and components, Progress would track completed tasks, tests and dependencies, and Experience would hold past fixes and error cases. In finance and compliance, Belief would cover applicable rules and evidence, Progress would track completed checks and open exceptions, and Experience would capture recurring issues and how they were previously resolved. Through what the team calls environment adapters, organizations can keep their own internal tools and state-management systems while adding BPE as a learnable coordination layer — without replacing their existing tooling. The research was posted to the preprint server arXiv (arXiv:2608.05446v1).

The finding matters beyond the benchmark itself. Until now, companies building agents for long, multi-step tasks have largely assumed they need to pay for a frontier model's API to get reliable results. This work suggests that a well-designed memory and tool-use layer can let a much smaller, cheaper open model close that gap, which could let more teams run agent workloads on their own infrastructure rather than routing every step through an expensive frontier API. The result comes from a single benchmark, ALFWorld, under controlled conditions, so translating it into production agent systems will still require further testing.