RecoveryBench v1¶
RecoveryBench asks a narrow mechanism question: when the starting checkpoint, teacher, task schedule, optimizer and update count are controlled, does scoring fresh states visited by the current student improve SQLite tool-error recovery over distillation on states frozen from the cold-start student?
It is not an alignment benchmark and it does not treat SFT and post-SFT OPD as interchangeable stages. SFT establishes task and protocol competence; OPD is a teacher-student mechanism whose transferred behavior depends on the teacher.
Result¶
The preregistered primary hypothesis was not supported. Under eight equal optimizer updates, frozen-student KD outperformed strict fresh-state OPD on both primary endpoints:
| method | strict success, mean (seed range) | recovery after error, mean |
|---|---|---|
| cold start | 10.7% (5.5-21.1%) | 13.6% |
| continued oracle SFT | 4.9% (0.0-9.4%) | 1.8% |
| oracle-state offline KD | 33.1% (0.0-54.7%) | 31.9% |
| frozen-student-state KD | 23.2% (9.4-36.7%) | 22.8% |
| strict fresh-state OPD | 10.9% (0.0-26.6%) | 9.1% |
| strict fresh-state OPD, 50% positions | 27.3% (10.9-49.2%) | 20.7% |
The task-paired difference, fresh OPD minus frozen-student KD, was -12.24 percentage points for strict success (95% paired bootstrap interval -15.89 to -8.59; 384 pairs) and -13.79 points for recovery after error (-20.69 to -6.90; 116 pairs where both trajectories had a structured tool error). These intervals describe paired task outcomes; three student seeds do not by themselves justify a broad significance or generalization claim.
Controlled design¶
- Student:
Qwen/Qwen3-0.6Batc1899de289a04d12100db370d81485cdf75e47ca. - Teacher base:
Qwen/Qwen3-1.7Bat70d244cc86ccca08cf5af4e1e306ecf908b1ad5e. - Qualified teacher adapter:
DaoyuanLi/mini-verl-qwen3-1.7b-sqlite-recovery-teacherat immutable revisioneb2747895ec32dab47c5b50c2d8aa9c0d9701e0d; adapter weights SHA-2565355f7007efb904d1b45a1aeb9b73b479b6f52025ab92502ab7895706155b2ba. - Seeds:
1234,20260727,20260801; fixed split seed20260801. - Tasks: 256 train, 96 eval and 128 final test tasks from 12 structurally disjoint templates.
- Same cold-start checkpoint and paired final-test task IDs within each seed.
- Primary: eight continuation updates. Secondary: 6,224 selected positions and a preregistered 50-second continuation target.
The selected teacher passed the independent eval-only gate at 90.6% strict success, 81.2% recovery after error, 100% parse-valid tool calls and 87.1% tool execution success. The historical calculator teacher failed the same gate. No test outcome was used to select the teacher.
Budget views and cost¶
The equal-selected-position view reached the 6,224 target at optimizer-step boundaries for all nine runs. Overshoot ranged from 0 to 646 positions. Because all three methods reached the boundary after eight steps, its quality outcomes match the primary view while independently preserving task-level execution evidence.
The nominal wall-time view needs a precise caveat. Fresh OPD crossed the 50-second continuation gate in one indivisible step (88.36-121.30 seconds on the internal continuation timer; 114.89-162.67 seconds for the complete train call). SFT and frozen KD instead completed the preregistered eight-cycle ceiling before their internal continuation timer crossed 50 seconds; their complete train calls averaged 51.92 and 51.63 seconds. This is a cycle-capped wall-time diagnostic, not an exact equal-time comparison. The completed artifact is preserved as run, per the steering instruction not to restart or reconfigure the frozen final experiment.
Under equal updates, strict fresh OPD averaged 686.80 continuation seconds, versus 52.10 seconds for frozen KD. The budget-50 position selector queried 49.77% of model-generated positions but averaged 720.76 seconds; selecting fewer positions did not reduce the required teacher backbone forwards and did not reduce wall time in this implementation.
Teacher preparation cost is reported separately: 2,310.75 seconds once, 462.15 seconds amortized over five students, or 231.07 seconds over ten. A 687.05-second diagnostic is excluded from that selected-teacher preparation total and remains recorded.
Interpretation¶
RecoveryBench isolates the value and cost of fresh student-visited states. On this Qwen3 pair, SQLite recovery task family and RTX 4080, fresh-state OPD did not justify its added cost. Oracle-state offline KD had the highest three-seed mean, while the 50%-position fresh arm was highly seed-variable. This result does not show that OPD is generally ineffective, that offline KD always wins, or that ordinary task success is the right endpoint for alignment.
The calculator benchmark remains a separate case study: teacher protocol qualification prevented collapse but only tied continued SFT on a saturated task. Together, the two studies motivate qualification and pilot evidence before choosing an online teacher-student mechanism.
Immutable artifacts¶
| artifact | SHA-256 |
|---|---|
| equal updates | 6ce2e6837e12b99ebc4fad6d27ce3e69c92e295ff3b9b60e0f68c2d308022384 |
| equal selected tokens | fe4c9afc799724dfe7a32e631676a1e5177c44559a7374d2ea31da135354f137 |
| wall-time diagnostic | 425b0fa568f37b09e61af731d3da5009bd3833bddde6efaf2c66e9dba8355cbe |
| compact task results | aff96bffc6da27240a852410ac041bd4d95badf34cad030e6f437be1491a55ad |
| paired analysis | 8a6891f74aed80f07ec00d5ea1909895c579346e1abbb1d5d95a354bb46c6b81 |
The three raw task-artifact sets contain 36 files and 4,608 trajectories. Their
individual hashes remain embedded in the schema-v3 result JSON. The
preregistration is commit 7087b3a333463b88a62ffed73daee2c85d039145, revision
1.3 digest 9c4c2ec19a56cebb2b2c1c0f3c7e504a9285467c99ae1590488251fbf2ff3934.
Reproduce the committed analysis and figures with:
python scripts/publish_recoverybench_artifacts.py \
--equal-updates benchmarks/results/recoverybench-v1-equal-updates.json \
--equal-selected-tokens benchmarks/results/recoverybench-v1-equal-selected-tokens.json \
--equal-wall-time benchmarks/results/recoverybench-v1-equal-wall-time.json
See the preregistration, machine-readable schema and technical report for the full method and limitations.