- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper introduces GALA, a cross-tokenizer distillation framework that combines teacher supervision with group-relative policy learning. It removes conflicting components from teacher gradients and controls the resulting updates through a learned gate and a relative norm cap. Experiments on Countdown report improvements over SFT and an unmodified aligned distillation baseline.
Critical strengths
None
Critical weaknesses
-
The empirical evaluation is too narrow, and comparisons with prior methods are insufficient. Evaluation on Countdown alone is insufficient to establish the broader effectiveness of the proposed method. Experiments on more diverse mathematical reasoning benchmarks, such as GSM8K, MATH, and AIME, are needed. Although the paper includes SFT, an externally reported GOLD result, and internal ablations, it does not directly evaluate the other cross-tokenizer distillation methods discussed in the related work. These omissions limit confidence in both the method's generality and its competitiveness with existing approaches.
-
The method is difficult to follow, and its contribution is unclear. The arithmetic verifier is tailored to Countdown, and the paper does not establish how the approach would extend to broader mathematical reasoning involving natural-language solutions. Moreover, the cross-tokenizer alignment is inherited from GOLD, leaving the contribution beyond this existing component insufficiently clear to me. The method section would benefit from clearer explanations of the notation and how the components in the equations fit together.
Questions for authors
- Can the authors provide additional results on more diverse mathematical reasoning benchmarks and comparisons with prior cross-tokenizer distillation methods to establish the effectiveness and generality of the proposed approach?