caiovicentino1/Eikos-27B-FP8-AgentRewardBench

Eikos-27B-FP8 as an AgentRewardBench judge

Eikos-27B-FP8 is an open-weight typed-decision model (MIT) that answers typed questions in one forward pass, with a probability for every option. This repository holds its judgments on AgentRewardBench (Lù et al., 2025), in the official judgment format, and the official scorer's output.

Results (test split, 1,106 trajectories, official scripts/score_judgments.py)

	Overall	AssistantBench	VisualWebArena	WebArena	WorkArena	WorkArena++
Precision	78.9	80.0	62.2	82.9	100.0	90.0

Overall recall is 73.6 and overall F1 76.1. For reference, the best published LLM judge, GPT-4o (Axtree), has 69.8 / 83.1 / 75.9. The expert-annotated success rate on the test split is 26.7%.

How it was run

• Input: the exact text the official text-only (Axtree) judges read, taken from the published gpt-4o-noscreen judgments: the user goal, every step (URL, action, the agent's own reasoning) and the last accessibility tree.
  • One context sentence goes before it, and the official final formatting instruction is dropped, because the questions are typed.
  • The judge gets the regular message if it fits a 62,000-token budget, otherwise pruned. This is the official fallback order.
• Questions: the four official questions, typed, answered together in one request:
  • success: yes/no, "Successful" / "Unsuccessful";
  • side effects: yes/no;
  • optimality: a 4-level score with the official labels;
  • looping: yes/no.
• Labels: a label is "Successful" / "Yes" when its probability is at least 0.5; optimality takes the most probable level. No examples, no prompt tuning and no threshold tuning.
• Serving: Eikos-27B-FP8 with its published serve_vllm.sh flags, except --max-model-len 65536, and serve.py on one RTX PRO 6000. Requests went one at a time.
• Scoring: the official scripts/score_judgments.py, unmodified. Only the judge list and display name were set.
  • The same setup first reproduced the published GPT-4o (Axtree) row (69.8 / 83.1 / 75.9) and the rule-based row (83.8 / 55.9 / 67.1).
• Protocol: fixed before the first judgment.
  • Four trajectories needed a fallback. Their inputs of 50k–61k tokens overflowed the 64k context after JSON escaping, and the server answered with an error. The runner first recorded those 4 as unanswered; they were then judged again on the pruned input, following the fallback rule.
  • With the 4 unanswered, the overall figures were 79.1 / 73.5 / 76.2.
• Each judgment file keeps the model's probabilities (probabilities), the input variant and its token count (input), and the latency (seconds).

Files

• results.json: the official scorer's results.json for this judge (human and judge labels for every benchmark, agent and category).
• metrics_test.csv: the official scorer's metrics for this judge on the test split.
• <benchmark>/<agent>/eikos-27b-fp8-noscreen/<task>.json: one judgment per trajectory, 1,302 in all (dev + test).

Terms of use

This repository contains data derived from AgentRewardBench, so its terms of use apply:

By downloading this Dataset, you agree to comply with the following terms of use:

• Restrictions: You agree not to use the Dataset in any way that is unlawful or would infringe upon the rights of others.
• Acknowledgment: By using the Dataset, you acknowledge that the Dataset may contain data derived from third-party sources, and you agree to abide by any additional terms and conditions that may apply to such third-party data.
• Fair Use Declaration: The Dataset may be used for research if it constitutes "fair use" under copyright laws within your jurisdiction. You are responsible for ensuring your use complies with applicable laws.

Derivatives must also include the terms of use above.

AgentRewardBench: Xing Han Lù et al., AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories, arXiv:2504.08942.