Hugging Face Releases Multi-Harness RL Guide and Open Tooling

Hugging Face's opening argument on X is a comparison: the same model, with the same weights, scores 62% in one agent harness and only 33% in another. This comparison is used to support the motivation for the entire multi-harness reinforcement learning guide—if the weights are identical yet the scores differ by nearly a factor of two, then benchmark results cannot be attributed to the model alone.

To be clear, this comparison comes from a single Hugging Face post; it is the publisher's own account, with no first-party paper or independent reproduction provided, and the post does not specify which model, which harness, or which evaluation tasks were involved. It is therefore better treated as an observation awaiting verification than as an established conclusion.

Commenter Ceyhun extends this, saying that a gap of 62 versus 33 under the same weights means half of every agent leaderboard is really a harness leaderboard. This too is the commenter's judgment, not a data conclusion from the publisher.

The key design of this approach is "don't modify the harness." The method is to point the harness at a proxy rather than directly at the model. The proxy supports the four API formats used by coding agents: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Gemini.

The proxy's role is to record the exact token ids and logprobs sampled by vLLM, and training is built on that data. According to the publisher, using this approach requires no changes to any line of code in Claude Code, Codex, or OpenCode.

This design decouples "training infrastructure" from the "agent front end": the harness stays as it is, and sampling and training happen at the proxy layer. The cost is that the approach depends on vLLM for sampling and recording token ids and logprobs, and the post does not state its applicability to other inference backends.

The main result the publisher gives is: after Liquid AI's LFM2.5-2.6B was trained on four harnesses simultaneously, its score rose from 42% to 54%.

As a comparison, training only in OpenCode raised OpenCode's score from 34% to 58%; but the publisher also says the multi-harness model improved across the board. Put together, these two numbers point to a question that has not yet been answered: single-harness training yields a larger gain on the target harness, while the advantage of multi-harness training lies in coverage.

Another result concerns efficiency: because a small reward is given for completing tasks in fewer steps, tool calls dropped by 31%. The post does not say whether this reduction comes at the cost of task success rate.

All of the above results are the publisher's own account and lack third-party verification; the post also gives no details on training compute, data scale, hyperparameters, or evaluation benchmarks.

The team also tried a more common path: fine-tuning on 3,189 successful rollouts from Qwen3.8-27B. The publisher says imitation learning stalled at 47.5%, below the results of the two reinforcement learning runs.

From this comparison one can infer that the team argues reinforcement learning is superior to imitation learning that copies a larger model—this is an interpretation drawn from the data comparison, not a causal conclusion the publisher directly states. Whether the gap between imitation learning's 47.5% and the reinforcement learning results is statistically significant is not stated in the post.

Hugging Face says everything is open source, including the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code, and all seven trained models.

However, the scope of open source is currently only asserted by the publisher; the post does not list specific links or licenses for the repositories or models, nor does it disclose the specific identity, scale, and training configuration of the seven open-source models. Whether the guide and toolchain can be independently reproduced by third parties remains unverified.

Stringing the above facts together, the question this line of work is trying to answer is not "how do we get a higher score on some harness," but "how should a model be trained when agents will run across dozens of harnesses." The publisher's judgment is: open models can now be trained for each harness, and anyone can do it.

Commenter Ceyhun offers a related perspective: cross-harness training is equivalent to domain randomization over tool formats and should become the default practice. This is the commenter's view, not the publisher's conclusion.

From the 62% versus 33% comparison one can infer that agent benchmark scores may be significantly affected by harness implementation, rather than being determined by model weights alone. But whether the gains from multi-harness training transfer to harnesses not involved in training is currently unclear.