Publish public agentic benchmark report and cleaned pass@5 results
Browse files- BENCHMARK-REPORT.md +25 -0
BENCHMARK-REPORT.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Benchmark report — public agentic RE environments
|
| 2 |
+
|
| 3 |
+
This report covers the public agentic family in `binary-reasoning-v1`: 29 flag-hunt reverse-engineering environments. Each environment ships a binary and asks for an input string that makes the program accept it. Grading is oracle-based: the harness runs the shipped binary with the model answer as `argv[1]`; pass means stdout contains `OK` and exit code is 0. The same 29 environments are packaged as Harbor task folders in `harbor_pilot_29.zip`; each folder includes a derivation-based reference solution script and explanation.
|
| 4 |
+
|
| 5 |
+
## Headline numbers
|
| 6 |
+
|
| 7 |
+
| run | scope | result | interval / caveats |
|
| 8 |
+
|---|---:|---:|---|
|
| 9 |
+
| Claude Opus 5.5, agentic pass@5 | all 29 public agentic envs; 5 attempts each = **145 completed runs** | **29/29 envs solved; 145/145 completed attempts correct** | Env-level Wilson 95% interval for pass@5: **88.3–100.0%**; attempt-level interval: **97.4–100.0%**. Three initial runs returned empty output and were re-run; counts are over the completed runs only. Non-isolated host run: prompts pointed to binaries in a source directory that also contained the stored-answer file, attempts shared host scratch folders, and only final messages were retained. Treat this as an oracle check and an upper bound, not a sealed-evaluation result. Byte-exact agreement with stored gold confirms the shipped binaries accepted those answers; it is not proof of uniqueness. |
|
| 10 |
+
|
| 11 |
+
## Open-weight pass@k status
|
| 12 |
+
|
| 13 |
+
No full open-weight leaderboard is published yet.
|
| 14 |
+
|
| 15 |
+
## Method notes
|
| 16 |
+
|
| 17 |
+
- Opus run: fresh session per completed attempt, Bash and file-read tools only, no web search. Three initial runs returned empty output and were re-run; counts are over the 145 completed runs. Solve cost median 43.8 s, mean 47.3 s, range 22.7–121.0 s, mean 8.4 agent turns. Cross-architecture environments solved every time but were slower: about 67.9 s mean for cross-arch versus about 43.0 s mean for x86/x86-PE once QEMU and normal RE tools were available.
|
| 18 |
+
- Harbor reward chain: all 29 environments were separately scored in real Docker containers: the reference solve script reproduces the gold answer → reward 1; wrong answer, stubbed binary, and both-binary-copies-stubbed controls → reward 0. That is an oracle/grading proof, not a difficulty or uniqueness claim.
|
| 19 |
+
|
| 20 |
+
## Honest reading
|
| 21 |
+
|
| 22 |
+
The public agentic tier is **warm-up level for a frontier agent**: Opus with tools solved every environment, so the public set is useful as a free demo, oracle check, and calibration starter, but it is not strong RL training material by itself. The paid product direction is pilot-sized private environments whose acceptance band is measured from the buyer's own reruns with our harness and graders; we supply analysis of those reruns.
|
| 23 |
+
|
| 24 |
+
Harbor package: `https://huggingface.co/datasets/S3RL4/binary-reasoning-v1/resolve/main/harbor_pilot_29.zip`
|
| 25 |
+
Dataset: `https://huggingface.co/datasets/S3RL4/binary-reasoning-v1`
|