Pith. sign in

SFT memorizes, RL generalizes: A comparative study of foundation model post-training

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it

citation-role summary

background 1

citation-polarity summary

fields

cs.LG 1

years

2025 1

verdicts

CONDITIONAL 1

roles

background 1

polarities

unclear 1

representative citing papers

RAST: Reasoning Activation in LLMs via Small-model Transfer

cs.LG · 2025-05-30 · conditional · novelty 5.0

Injecting the logit gap between a small RL-trained model and its base into a larger base model at decoding time improves reasoning accuracy on math and code benchmarks, sometimes matching or exceeding RL training of the large model.

citing papers explorer

Showing 1 of 1 citing paper.

  • RAST: Reasoning Activation in LLMs via Small-model Transfer cs.LG · 2025-05-30 · conditional · none · ref 4

    Injecting the logit gap between a small RL-trained model and its base into a larger base model at decoding time improves reasoning accuracy on math and code benchmarks, sometimes matching or exceeding RL training of the large model.