Research map / Reinforcement learning / RL from feedback
Language Model Alignment with Elastic Reset
Michael Noukhovitch, Samuel Lavoie, Florian Strub, Aaron Courville
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 Elastic Reset to REINFORCE | Translation Game · Translation Game validation | FR→EN BLEU: 29.6±0.3 -> 36.3±0.1 | Improved |
| Add Elastic Reset to PPO | IMDB mock sentiment · IMDB mock sentiment test | Sentiment: .596±0.02 -> .611±0.02 | Improved |
| Add Elastic Reset to PPO | StackLLaMA · StackLLaMA validation | Perplexity: 4.62 -> 4.57 | Improved |
| Add Elastic Reset to PPO | StackLLaMA · HumanEval | pass@1: 7.8 -> 11.0 | Improved |
| Add the reset of the EMA to the initial model (ablation) | Reset to EMA on IMDB · IMDB mock sentiment | reward–drift tradeoff | Improved |
| Reset the policy to the initial model instead of the EMA (ablation) | Elastic Reset on IMDB · IMDB mock sentiment | reward–drift tradeoff | Mixed |
| Use an EMA policy reset without resetting the EMA to the initial model (ablation) | Elastic Reset on IMDB · IMDB mock sentiment | reward–drift tradeoff | Mixed |
| Vary the EMA decay rate (ablation) | Elastic Reset · IMDB mock sentiment | reward–drift tradeoff | No clear effect |
| Vary the KL coefficient (ablation) | Elastic Reset · IMDB mock sentiment | reward–drift tradeoff | No clear effect |
| Remove the KL penalty (ablation) | Elastic Reset · IMDB mock sentiment | reward–drift tradeoff | No clear effect |
| Vary the reset frequency (ablation) | Elastic Reset · IMDB mock sentiment | reward–drift tradeoff | No clear effect |
| Keep an EMA model without resetting the policy to it (ablation) | PPO on IMDB · IMDB mock sentiment | reward–drift tradeoff | Improved |
Closest papers by meaning
- Theoretical Limits of Language Model Alignment (NeurIPS 2026 · similarity 0.91)
- Decoding-time Realignment of Language Models (ICML 2024 · similarity 0.90)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.90)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.90)
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.90)
- Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback (AAAI 2025 · similarity 0.90)
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.
