Research map / Large language models / Instruction and fine-tuning
Don't Force the Fit: Bounded Log-Likelihood Loss for Enhanced Reasoning in Large Language Models
Feng Zhao, Hong Zhang, Yu Yang, Ruilin Zhao, Guandong Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace cross-entropy with bounded log-likelihood loss log(2 - p) | SFT on NuminaMath-CoT 100k · 7 math benchmarks, Qwen2.5-Math-1.5B | Overall Avg. accuracy: 28.36 -> 41.56 | Improved |
| standard SFT on Qwen3-8B-Base | Qwen3-8B-Base · 7 math benchmarks | Overall Avg. accuracy: 38.74 -> 35.24 | Hurt |
| mask loss of tokens with probability below threshold (5%, 10%, 15% quantile) (ablation) | SFT Qwen2.5-Math-1.5B · AIME, Minerva, Olympiad, AMC, MATH500, GSM8K | accuracy | Improved |
| lower learning rate 1e-6 for SFT (ablation) | SFT Qwen2.5-Math-1.5B · 5 math benchmarks avg | Avg. accuracy: 31.72 -> 23.43 | Hurt |
| add gradient clipping (0.3 or 0.5) to SFT (ablation) | SFT Qwen2.5-Math-1.5B · 5 math benchmarks avg | Avg. accuracy: 31.72 -> 32.09 (0.3), 32.13 (0.5) | No clear effect |
| train SFT for 2 epochs (ablation) | SFT Qwen2.5-Math-7B · 5 math benchmarks avg | Avg. accuracy: 40.72 -> 43.60 (BLL 51.80) | Improved |
| reduce beta toward 0 in beta-balanced loss (complement suppression only) (ablation) | BLL-Loss · GSM8K, MATH500, Olympiad, Minerva, AMC | accuracy | Improved |
| BLL-Loss versus SFT at different training set sizes (10k-100k) (ablation) | Qwen2.5-Math-1.5B · data scaling | accuracy | Improved |
| use BLL-Loss for code SFT (oss instruct) | Qwen3-4B-Base · MBPP | accuracy: 38.40 -> 61.80 | Improved |
| use BLL-Loss for general reasoning SFT (alpaca-gpt4) | Qwen3-4B-Base · BBH | accuracy: 51.49 -> 65.85 | Improved |
| use GEM loss instead of SFT | Qwen2.5-Math-1.5B · 7 math benchmarks | Overall Avg. accuracy: 18.79 (Base) -> 14.04 | Hurt |
| add NEFT noise to SFT | SFT Qwen2.5-Math-7B · 7 math benchmarks | Overall Avg. accuracy: 35.95 -> 35.23 | No clear effect |
Closest papers by meaning
- UFT: Unifying Supervised and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models (NeurIPS 2025 · similarity 0.90)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.89)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
