feat(bench): HumanEval steering adapter + off-box runExperiment (gate: steering null)#197
Merged
Merged
Conversation
… result Wires the observe→steer EFFICACY A/B (the experiment humaneval-gate.mts named as "the next one": a real runLoop rollout that self-corrects across rounds, vs the gate's stateless selection): - benchmarks/humaneval.ts: HumanEval BenchmarkAdapter — loader + Docker deployable verifier MOVED here from humaneval-gate.mts (one home, no duplication; the gate now imports them). judge = run candidate against the task's own tests in an isolated --network=none container. - rsi.ts: BACKEND=router runs the worker OFF-BOX (router chat-completion as the leaf executor via inlineSandboxClient(createExecutor)) — the real kernel + analyst steering with no sandbox dependency (sandbox→router egress is blocked, #984). Generalizes the canonical caller for every adapter, not just this run. Arm labels aligned to corpus-report's random*/refine* contract. Gate result (gpt-3.5-turbo, HumanEval hard half tasks 82-163, n=82, equal k=2, paired bootstrap B=10000): more-compute random@2 − blind@1 +12.2pp CI [+6.1, +19.5] SIGNIFICANT observe→steer refineAudit@2 − random@2 -1.2pp CI [-8.5, +6.1] n.s. generic-steer refinePush@2 − random@2 +1.2pp CI [-4.9, +7.3] n.s. At equal compute, observe→steer does NOT beat blind resampling on this deployable-checker domain; compute itself significantly does. A clean null for within-run steering here — consistent with the FinSearchComp null and the selection≠RSI audit. Drew's cheap models can't open the band off-box (kimi/glm unserved via router; deepseek-chat saturates at ~90%), hence gpt-3.5-turbo.
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.
What
Runs the observe→steer efficacy gate — the experiment
humaneval-gate.mtsitself names as "the next one": a realrunLooprollout that self-corrects across rounds, vs the gate's stateless selection. Ships the harness to run it on any adapter off-box.bench/src/benchmarks/humaneval.ts— HumanEvalBenchmarkAdapter. The loader + Docker deployable verifier are moved here fromhumaneval-gate.mts(one home, no duplication — the gate now imports them). Judge runs the candidate against the task's own tests in an isolated--network=nonecontainer.bench/src/rsi.ts—BACKEND=routerruns the worker off-box (router chat-completion as the leaf executor viainlineSandboxClient(createExecutor(...))): the real kernel + analyst steering, no sandbox dependency. Routes around the sandbox→router egress block (#984) and generalizes the canonical caller for every adapter. Arm labels aligned tocorpus-report'srandom*/refine*contract so the paired-bootstrap + BH report auto-detects the contrasts.adapters.ts— registershumaneval.Gate result
gpt-3.5-turbo, HumanEval hard half (tasks 82–163), n=82, equal k=2, paired bootstrap B=10000:
random@2 − blind@1)refineAudit@2 − random@2)refinePush@2 − random@2)At equal compute, observe→steer does NOT beat blind resampling on this deployable-checker domain; compute itself significantly does. A clean within-run-steering null — consistent with the FinSearchComp null and the selection≠RSI audit. The lone positive within-run signal remains depth>breadth on EOPS (agentic, multi-turn).
Notes
K=2because at ~55% pass@1,random@4≈96% would erase the contrast.BENCH=humaneval BACKEND=router WORKER_MODEL=gpt-3.5-turbo N=82 ROUNDS=2 OFFSET=82 tsx src/rsi.ts→tsx src/corpus-report.mts <corpus>.Test
pnpmtypecheck clean (bench tsconfig); the gate (refactored to import the moved primitives) ran live and produced the 54.9% band probe; the full A/B ran n=82 with 0 infra errors.