Research map / Large language models / Language, safety and interpretability
PonderLM: Pretraining Language Models to Ponder in Continuous Space
Boyi Zeng, Shixiang Song, Siyuan Huang, Yixuan Wang, He Li, Ziwei He, Xinbing Wang, Zhiyu li, Zhouhan Lin
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 |
|---|---|---|---|
| add pondering (3 steps) during pretraining | Pythia-2.8B · 9 downstream tasks 5-shot | average accuracy: 57.6 -> 61.5 | Improved |
| add pondering (3 steps) during pretraining | Pythia-1B · 9 downstream tasks 0-shot | average accuracy: 50.6 -> 55.0 | Improved |
| add pondering to GPT-2 and LLaMA pretraining | GPT-2 and LLaMA 405M-1.4B · Pile subset | pretraining loss | Improved |
| replace pondering embedding with last hidden state (Coconut-style) (ablation) | PonderPythia-70M · Pile validation | perplexity: 14.16 -> 15.64 | Hurt |
| replace pondering with looped transformer (3 loops) (ablation) | PonderPythia-70M · Pile validation | perplexity: 14.16 -> 15.33 | Hurt |
| replace pondering with pause tokens (3 pauses) (ablation) | PonderPythia-70M · Pile validation | perplexity: 14.16 -> 16.53 | Hurt |
| add linear projector on last hidden state (ablation) | Last Hidden State baseline · Pile validation | perplexity: 15.64 -> 15.64 | No clear effect |
| increase pondering steps from 3 to 6 (ablation) | PonderPythia-70M · Pile validation | perplexity: 14.16 -> 13.56 | Improved |
| increase top-K from 10 to 100 (ablation) | PonderPythia-70M · Pile validation | perplexity: 15.18 -> 14.21 | Improved |
| use full vocabulary instead of top-100 tokens (ablation) | PonderPythia-70M · Pile validation | perplexity: 14.21 -> 14.20 | No clear effect |
| train with randomized pondering steps in [1,10] (ablation) | PonderPythia-70M · Pile validation | loss | Improved |
| continually pretrain Pythia-1B with pondering instead of vanilla | Pythia-1B · 9 downstream tasks 5-shot | average accuracy: 50.7 -> 51.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Pythia-12B — Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling (ICML 2023) · better on 0 of 2 shared results
- Bloom-560M — BLOOM: A 176B-Parameter Open-Access Multilingual Language Model (JMLR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PonderLM-2: Pretraining LLM with Latent Thoughts in Continuous Space (ICML 2026 · similarity 0.91)
- Think before you speak: Training Language Models With Pause Tokens (ICLR 2024 · similarity 0.87)
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.86)
- TSLM: Tree-Structured Language Modeling for Divergent Thinking (ICLR 2026 · similarity 0.86)
- Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space (NeurIPS 2025 · similarity 0.85)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.85)
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.
