- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper asks whether language agents trained separately on math, code, and science can be merged into a model that performs better than the single-domain models. The authors find that merging does not always help: code and math usually benefit each other, whereas math and science are prone to negative transfer.
To address this, the authors propose Update-aware SAM (USA). Specifically, each domain first runs 10 steps of on-policy distillation (OPD), and the importance of each parameter is determined by the magnitude of its update. Parameters with larger updates are then assigned a larger SAM perturbation radius. Subsequent training requires the model to keep a low loss within this non-uniform perturbation neighborhood, with the goal of making these parameters more robust to the displacement introduced by merging. Each domain is still trained independently, and the final model is obtained by plain weight averaging, without sharing data or gradients across domains.
USA achieves strong results on all six transfer directions (across the math, code, and science domains) for both students (Qwen3-1.7B and Qwen3-4B), with average improvements of 4.55 and 4.01 percentage points over single-domain OPD, respectively.
Critical strengths
-
The problem being addressed is valuable. Compared with methods that require repeated joint multi-domain retraining, USA preserves the modular workflow of training each expert separately, while making the experts more suitable for merging already at training time.
-
The empirical results are strong and convincing. USA is effective not only on domain pairs that naturally exhibit positive transfer, but also improves the originally conflicting math–science pair. For example, for the 1.7B student on Math → Science, single-domain OPD reaches 48.42, plain weight averaging drops to 45.44, while USA reaches 53.50. For the 4B student in the same direction, the three numbers are 61.71, 59.02, and 67.04, respectively.
-
The experimental coverage and reporting are generally thorough. The main experiments and the code-interpreter experiments are run with five independent seeds, and both the main table and the per-benchmark tables report uncertainty. The appendix further extends the setting to up to five source domains and evaluates TIES, DARE, and AdaMerging. The search-agent experiments use different teacher scales and tool environments, and USA outperforms the compared baselines on all six directions without re-tuning its hyperparameters. These extensions show that the conclusions do not hinge on a single domain pair or a single merging operator.
-
The method is simple to implement, and the actual training overhead is reported. The method itself is concise (the two forward/backward passes of USA reuse the same batch of student trajectories and teacher scores, without requiring an additional full rollout), and it is well supported by the experiments.
Critical weaknesses
The scope of the evidence for broader generalization and multi-domain capability retention remains somewhat limited.
Both the students and the teachers in the main experiments come from the Qwen3 family, and the main ablations and mechanistic analyses focus on the 1.7B student. It is currently unclear whether the update-magnitude-driven scaling transfers equally well to other backbones.
Questions for authors
see weakness