Add OSWorld-V2 hybrid GUI+CLI evaluation experiment - #11
Merged
Conversation
Harness ablation: drive OSWorld-V2 (108 tasks) with a hybrid GUI+CLI agent (codex CLI injected in-VM + GUI channel) on the GPT-5.5 backbone, scored with OSWorld's native env.evaluate(). Same backbone gains +5.5 pt Binary (13.0->18.5%) at ~2x tool-call efficiency (149.8->77.5/task) vs the official batched row. Includes core code, launchers, and aggregated per-task results. Co-Authored-By: Claude <noreply@anthropic.com>
…ble 3) env.evaluate() is checkpoint-based with bounded model judgment, not a pure static check; align the README with the paper's description and drop the misleading agent-judge contrast. Co-Authored-By: Claude <noreply@anthropic.com>
Partial in OSWorld-V2 / the paper is the mean partial-credit score over all tasks, not the fraction of tasks scoring 0<s<1. Corrected the headline and analysis tables (hybrid 59.3 -> 49.6; de-infra 51.4) and the aggregation script. Binary and tool-call numbers unchanged. Co-Authored-By: Claude <noreply@anthropic.com>
Drop the cross-harness comparison rows (pure-CLI / claude / cuaclaw / official GUI) for now; keep only the codex hybrid row against the paper's official GPT-5.5 claim. Other harnesses to be added later. Co-Authored-By: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
A harness ablation on OSWorld-V2 (108 tasks): drive each task with a hybrid GUI+CLI agent — the OpenAI
codexCLI injected inside the VM (persistent state via shell/code) plus a GUI action channel — on the same GPT-5.5 backbone the paper benchmarks. Scored with OSWorld-V2's own evaluator (env.evaluate(), checkpoint-based with bounded model judgment) — the same grader as the paper's Table 3, so the numbers are directly comparable to the official GPT-5.5 row.(Partial = mean partial-credit score over all tasks, per the paper's Table 3.)
Same backbone, swapping the official batched loop for the codex hybrid harness lifts GPT-5.5 +5.5 pt Binary (13.0→18.5%) at ~2× tool-call efficiency, while partial credit holds steady (49.5→49.6) — the gains come from pushing near-complete tasks over the line, not broad partial progress. De-infra cohort (drop 4 infra-failure tasks): 104 tasks, 51.39% avg / 19.23% Binary.
Contents
experiments/osworld_v2_hybrid/: runner, codex/openclaw/claudecode agents, launchers, aggregated per-task results +RESULT_ANALYSIS.md.Notes
Cost/taskandOut tok/taskare intentionally—: the codex CLI logs only a combined token total (no output-only count or price).🤖 Generated with Claude Code