- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper identifies sensitivity to noisy reference and overcorrection as major limitations of SFT and claims that existing token reweighting methods are limited in addressing these issues. The authors introduce SCALE, a method that uses module-specific gating to adjust the strength of learned adaptations. The gates are trained by minimizing the model’s predictive entropy on the training data. Experiments QWEN model variants show improvements on mathematical reasoning and code generation.
Critical strengths
-
The author’s claims on the limitations of SFT are well-grounded.
-
The author’s claims on the limitations of existing token reweighting methods are convincing.
-
The authors evaluate on a variety of math and coding benchmarks.
Critical weaknesses
-
The motivation for using predictive entropy in adaptation-strength-control is unclear. The assumptions on entropy-task consistency are broad and experimentally unsupported.
-
The main experiments favor SCALE by design. Table1 and Table2 compare baselines (stage 1 only) and SCALE (stage 1 AND stage 2). Table 6 in the appendix addresses this concern only partially, but shows results on a single QWEN model.
-
The claimed benefits of SCALE are insufficiently demonstrated. It is unclear whether the method is enabling the model to actively “reverse” the learned feature as a result of SCALE training.
Questions for authors
-
How does SCALE differ from existing entropy-based test-time-training methods?
-
Can you provide experiment results conducted based on equal compute and data?
-
Can you provide results for other model families?
3 comments
I appreciate the response by the authors. However, the concerns raised in the critical weaknesses section still remain unanswered. In particular, the motivational experiment in Figure 1(left), which summarizes the central claims of the paper, requires more explanation and deeper analysis to be more convincing. I will keep my current score.
Thanks.
I would like to thank the reviewers for answering some of my concerns. I'll maintain my score.
Q1. How does SCALE differ from entropy-based test-time training?
Thank you for raising this distinction. SCALE trains its gates on demonstration prefixes before deployment. At inference, one forward pass gives the gate for each token. TENT updates its parameters on test inputs.
Q2. What does matching the data control, and how much compute does SCALE use?
With the same 20K additional math examples, SCALE [−4,4] gained 21.57 points over LoRA +20k in Table 6.
With the same Stage-I adapter, 20K prefixes, entropy objective and 157 updates, LoRA +20k w/ Entropy scored 28.48 and SCALE [−4,4] scored 37.84. With equal training time on one H200 (about 22.7 minutes), SCALE [−3,3] scored 34.80 versus 29.04 for continued LoRA.
Table 6 · Qwen2.5-Math-1.5B · Math Avg@16 (%)
Figure 4 · single-H200 training time (minutes)
Q3. Does the result transfer to another model family?
On Gemma-3-4B-PT, SCALE raised the five-benchmark average from 4.74% with LoRA to 7.98%.
Gemma-3-4B-PT · Math Avg@16 (%)