Research map / Probabilistic and scientific modeling / Variational and generative sampling
Variational Reasoning for Language Models
Xiangxin Zhou, Zichen Liu, Haonan Wang, Chao Du, Min Lin, Chongxuan Li, Liang Wang, Tianyu Pang
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 |
|---|---|---|---|
| condition the variational posterior on an answer hint y' (ablation) | Variational reasoning, Qwen3-4B-Base · math suite | average accuracy: 48.18 -> 55.72 | Improved |
| condition the variational posterior on an answer hint y' (ablation) | Variational reasoning, Qwen3-4B-Base · other suite | average accuracy: 37.80 -> 46.12 | Improved |
| use accuracy-based estimator instead of naive likelihood for pi(Y|x,z) (ablation) | Variational reasoning, Qwen2.5-7B-Instruct 1K · math suite | average accuracy: 43.01 -> 45.41 | Improved |
| use geometric-mean token likelihood instead of naive likelihood (ablation) | Variational reasoning, Qwen2.5-7B-Instruct 1K · other suite | average accuracy: 37.35 -> 39.42 | Improved |
| accuracy-based vs GML estimator (ablation) | Variational reasoning · four model scales | average accuracy: 55.72 vs 55.23 (Qwen3-4B math) | No clear effect |
| increase number of sampled traces K from 1 to 32 (ablation) | Variational reasoning, 1K data · reasoning benchmarks | accuracy | Improved |
| variational reasoning vs distillation on same data | Qwen3-4B-Base · math suite | average accuracy: 51.35 -> 55.72 | Improved |
| variational reasoning vs distillation on same data | Qwen3-8B-Base · other suite | average accuracy: 49.46 -> 54.69 | Improved |
| variational reasoning vs RLT | Qwen2.5-32B-Instruct · math suite | average accuracy: 70.43 -> 72.01 | Improved |
| switch prompt template A to B for the posterior | Variational reasoning · reasoning benchmarks | average accuracy | No clear effect |
| add variational weighting to SFT | Qwen3-4B/8B-Base · training | loss and gradient norm | Improved |
Closest papers by meaning
- Coupled Variational Reinforcement Learning for Language Model General Reasoning (ICML 2026 · similarity 0.91)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.88)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.88)
- Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.88)
- Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.88)
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.
