AI

Rethinking Token Reweighting for SFT: Suppress, Reverse, and Extrapolate Learned Features

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 11:18:24 UTC.

Briefings

This briefing presents findings from a 2026 study on token reweighting for supervised fine-tuning (SFT). The research argues that standard methods cannot reverse harmful learned features and that larger training weights do not equal feature extrapolation. Instead, the authors propose SCALE, an entropy-guided method that freezes pretrained models and SFT deltas to learn bounded gates for suppressing, reversing, or extrapolating learned features. Experiments on Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, and Qwen3-4B-Base show SCALE outperforms baselines in mathematical reasoning while maintaining competitive performance on general retention and code generation benchmarks.

Key points

• Standard token-reweighting methods assign nonnegative coefficients to demonstrated tokens, which allows them to suppress or amplify supervised updates but cannot reverse harmful features once learned.
• SCALE achieves mathematical-reasoning averages of 37.84, 43.60, and 36.57 on Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, and Qwen3-4B-Base respectively, exceeding corresponding baselines while remaining competitive on general-retention benchmarks.

Context and limitations

Results are limited to three specific models; generalization to other architectures remains unverified. The study uses preprint status and does not include causal claims or author opinions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 06/10/2026, 05:18:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibi3o4se5vmvxhx2ndd34dlpvmqs6m3adxbjmkxwuadiu56syg2wu