Loading
Life sciences · Preprint
arXiv · August 19, 2026
Early or partial results. Treat as a signal, not a conclusion.
This preprint introduces Group-Calibrated On-Policy Distillation (GC-OPD), a training method intended to improve language model performance on long-context reasoning tasks by reconciling token-level teacher guidance with task-level verifier rewards. Reported improvements on five benchmarks are numerically substantial but have not undergone peer review, external validation, or assessment of real-world utility.
Algorithmic method comparison on benchmarks with controlled ablations. Qwen3-4B and Qwen3-8B language models on five long-context evidence-aggregation tasks.. Intervention: Group-Calibrated On-Policy Distillation (GC-OPD) with relative-advantage-based credit assignment (RACA).. Compared with: Vanilla on-policy distillation (OPD) and baseline model performance..
Qwen3-4B five-benchmark average improved from 29.08 to 40.47 with GC-OPD, compared to 39.31 with vanilla OPD. Qwen3-8B five-benchmark average improved from 35.12 to 44.65 with GC-OPD, compared to 43.56 with vanilla OPD. Signed residual calibration outperformed alternative approaches in ablation studies.
Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.
Not applicable. This is a computational method paper with no clinical or medical domain application described.
This is an unrefereed preprint describing an algorithmic method for improving language model training on long-context tasks, with reported benchmark improvements but no peer review, external validation, or clinical/real-world outcome data.
As stated by the source record.
Quoted from the source exactly as published.
Not applicable. This is a computational method paper with no clinical or medical domain application described.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
On-policy distillation (OPD) trains a student on its own responses using dense token-level guidance from a stronger teacher. In long-context tasks, however, token-level teacher support can favor locally plausible responses that omit evidence distributed across the input or violate global task constraints. Task-specific verifiers, in contrast, evaluate task completion at the response level and may return graded rewards that reflect partial success. We diagnose this mismatch on fixed responses from two representative long-context evidence-aggregation tasks. Across longer input ranges, trajectory-level OPD scores become progressively less aligned with verifier rewards, indicating teacher-verifier disagreement. Motivated by this observation, we introduce Group-Calibrated On-Policy Distillation (GC-OPD). GC-OPD separately normalizes verifier rewards and trajectory-level OPD scores within each rollout group and uses their difference as a signed teacher-verifier disagreement residual. Relative-advantage-based credit assignment (RACA) distributes this trajectory-level residual across tokens according to their relative OPD advantages while preserving the original OPD signal. Across five long-context benchmarks, post-training with GC-OPD raises the five-benchmark averages of the official Qwen3-4B and Qwen3-8B checkpoints from 29.08 to 40.47 and from 35.12 to 44.65, respectively. Vanilla OPD reaches 39.31 and 43.56 under the same setup. Controlled ablations show that the signed residual is more effective than either an additional OPD-derived term or direct group-normalized verifier reward addition, while RACA further improves over uniform token allocation. Together, these results demonstrate that group-relative residual calibration can incorporate verifier outcomes without discarding dense token-level guidance. Code is available at https://github.com/SolereZhang/GC-OPD.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.