Skip to content

Latest commit

 

History

History
71 lines (46 loc) · 3.27 KB

File metadata and controls

71 lines (46 loc) · 3.27 KB

Who&When Localization Benchmark

Headline

This is an honest coverage-limited benchmark of agent-bisect against Who&When labels. It is not an apples-to-apples comparison to the published Who&When LLM attribution baseline.

metric result
exact-step accuracy 0/181 (0.000)
cascade-membership accuracy 0/181 (0.000)
coverage-gap rate 181/181 (1.000)
rows processed 184
included labels 181
excluded labels 3

Confidence Split

confidence exact correct total exact accuracy share of labels
HIGH 0 0 NA 0.000
LOW 0 0 NA 0.000
NA 0 181 0.000 1.000

By Config

config exact cascade coverage gaps labels
Algorithm-Generated 0/126 (0.000) 0/126 (0.000) 126/126 (1.000) 126
Hand-Crafted 0/55 (0.000) 0/55 (0.000) 55/55 (1.000) 55

Excluded Labels

reason count
mistake_step_out_of_range 3

What Was Scored

Dataset: Kevin355/Who_and_When at revision 59b9fcba1aaed7bbf206b5f4d3c68b8face2f49c.

Manifest hash: 84bb9d16f34f36d16e145854aeb27c65d7a7a732848ea5a0bcd8933959c8e2af.

Who&When labels mistake_agent and mistake_step over multi-agent conversation histories. This benchmark maps each history item one-to-one into an Activity, with mistake_step treated as a zero-based Activity.step_index.

The adapter does not infer hidden file edits, commands, exit codes, or test failures from free text. Agent messages are preserved as llm_call/user_msg; terminal/computer records without recoverable commands become explicit unmapped activities.

Adapter Coverage

activity kind count share
llm_call 3692 0.902
unmapped 342 0.084
user_msg 58 0.014

Rows: 184; activities: 4092; unmapped activities: 342.

Baseline Context

The Who&When paper reports a best semantic step-level attribution result of about 14.2% for GPT-4o Step-by-Step judging, from the cells 25.51%, 7.02%, 15.31%, 8.77%. That method judges multi-agent natural-language logs and sometimes has final-answer ground truth. agent-bisect instead localizes deterministic gate-visible breaks in normalized transcripts.

Because the Who&When histories do not expose deterministic file/test gate failures in the format agent-bisect requires, the clean comparison is ill-posed. The result above is a standalone visibility result over the full labeled set, not an Nx better claim.

Lineage And License

Sources: Who&When on Hugging Face, paper, and GitHub repository.

Raw Who&When rows are fetched from source into ignored local data/ only. They are not committed. The Hugging Face dataset card does not declare a dataset license; the GitHub repository is MIT-licensed, and the dataset is based on GAIA and AssistantBench tasks, so this repository ships only code, synthetic tests, and aggregate results.

Determinism

The scorer sorts labels and predictions deterministically and writes canonical JSON to reports/who-when-benchmark.json. Re-running over the same manifest should produce byte-identical JSON.