# Language Model Alignment with Elastic Reset

arXiv: https://arxiv.org/abs/2312.07551 · HTML: https://mlsanity.com/paper/2312.07551/

## Key facts

- Cited by 34 papers in the map (data through 14 Oct 2026).
- Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 39 of its 67 references that are in the map.
- First posted on arXiv 6 Dec 2023.

## Placed in

- LLM alignment / RLHF / preference optimisation / [LLM preference optimization methods](https://mlsanity.com/alignment/llm-preference-optimization-methods/)

## Builds on (most cited first, 30 of 39)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [Evaluating Large Language Models Trained on Code](https://mlsanity.com/paper/2107.03374/) (2021)
- [Emerging Properties in Self-Supervised Vision Transformers](https://mlsanity.com/paper/2104.14294/) (2021)
- [Overcoming catastrophic forgetting in neural networks](https://mlsanity.com/paper/1612.00796/) (2016)
- [Neural Machine Translation by Jointly Learning to Align and Translate](https://mlsanity.com/paper/1409.0473/) (2014)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [Deep reinforcement learning from human preferences](https://mlsanity.com/paper/1706.03741/) (2017)
- [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2204.05862/) (2022)
- [Sequence to Sequence Learning with Neural Networks](https://mlsanity.com/paper/1409.3215/) (2014)
- [Categorical Reparameterization with Gumbel-Softmax](https://mlsanity.com/paper/1611.01144/) (2016)
- [Learning to summarize from human feedback](https://mlsanity.com/paper/2009.01325/) (2020)
- [Trust Region Policy Optimization](https://mlsanity.com/paper/1502.05477/) (2015)
- [Fine-Tuning Language Models from Human Preferences](https://mlsanity.com/paper/1909.08593/) (2019)
- [A Call for Clarity in Reporting BLEU Scores](https://mlsanity.com/paper/1804.08771/) (2018)
- [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://mlsanity.com/paper/1904.01038/) (2019)
- [The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables](https://mlsanity.com/paper/1611.00712/) (2016)
- [Scaling Laws for Reward Model Overoptimization](https://mlsanity.com/paper/2210.10760/) (2022)
- [A General Language Assistant as a Laboratory for Alignment](https://mlsanity.com/paper/2112.00861/) (2021)
- [LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale](https://mlsanity.com/paper/2208.07339/) (2022)
- [Datasets: A Community Library for Natural Language Processing](https://mlsanity.com/paper/2109.02846/) (2021)
- [Way Off-Policy Batch Deep Reinforcement Learning of Implicit Human Preferences in Dialog](https://mlsanity.com/paper/1907.00456/) (2019)
- [Multi30K: Multilingual English-German Image Descriptions](https://mlsanity.com/paper/1605.00459/) (2016)
- [The Primacy Bias in Deep Reinforcement Learning](https://mlsanity.com/paper/2205.07802/) (2022)
- [Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization](https://mlsanity.com/paper/2210.01241/) (2022)
- [Deal or No Deal? End-to-End Learning for Negotiation Dialogues](https://mlsanity.com/paper/1706.05125/) (2017)
- [Emergent Multi-Agent Communication in the Deep Learning Era](https://mlsanity.com/paper/2006.02419/) (2020)

## Built on by (most cited first, 30 of 34)

- [WARM: On the Benefits of Weight Averaged Reward Models](https://mlsanity.com/paper/2401.12187/) (2024)
- [A Survey on Self-Evolution of Large Language Models](https://mlsanity.com/paper/2404.14387/) (2024)
- [RLHF Deciphered: A Critical Analysis of Reinforcement Learning from Human Feedback for LLMs](https://mlsanity.com/paper/2404.08555/) (2024)
- [Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Models](https://mlsanity.com/paper/2410.18252/) (2024)
- [SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection](https://mlsanity.com/paper/2410.07471/) (2024)
- [Self-Exploring Language Models: Active Preference Elicitation for Online Alignment](https://mlsanity.com/paper/2405.19332/) (2024)
- [The N+ Implementation Details of RLHF with PPO: A Case Study on TL;DR Summarization](https://mlsanity.com/paper/2403.17031/) (2024)
- [WARP: On the Benefits of Weight Averaged Rewarded Policies](https://mlsanity.com/paper/2406.16768/) (2024)
- [Coevolving with the Other You: Fine-Tuning LLM with Sequential Cooperative Multi-Agent Reinforcement Learning](https://mlsanity.com/paper/2410.06101/) (2024)
- [Online Merging Optimizers for Boosting Rewards and Mitigating Tax in Alignment](https://mlsanity.com/paper/2405.17931/) (2024)
- [Learning to Summarize from LLM-generated Feedback](https://mlsanity.com/paper/2410.13116/) (2024)
- [Model Surgery: Modulating LLM's Behavior Via Simple Parameter Editing](https://mlsanity.com/paper/2407.08770/) (2024)
- [Countering Reward Over-optimization in LLM with Demonstration-Guided Reinforcement Learning](https://mlsanity.com/paper/2404.19409/) (2024)
- [The Dark Side of Function Calling: Pathways to Jailbreaking Large Language Models](https://mlsanity.com/paper/2407.17915/) (2024)
- [Neuroplastic Expansion in Deep Reinforcement Learning](https://mlsanity.com/paper/2410.07994/) (2024)
- [Diffusion Guided Language Modeling](https://mlsanity.com/paper/2408.04220/) (2024)
- [Reward Difference Optimization For Sample Reweighting In Offline RLHF](https://mlsanity.com/paper/2408.09385/) (2024)
- [Reward-Augmented Data Enhances Direct Preference Alignment of LLMs](https://mlsanity.com/paper/2410.08067/) (2024)
- [Avoiding mathbfexp(Rmax) scaling in RLHF through Preference-based Exploration](https://mlsanity.com/paper/2502.00666/) (2025)
- [Gradient Regularization Mitigates Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards](https://mlsanity.com/paper/2602.18037/) (2026)
- [Understanding Likelihood Over-optimisation in Direct Alignment Algorithms](https://mlsanity.com/paper/2410.11677/) (2024)
- [Pretrain Value, Not Reward: Decoupled Value Policy Optimization](https://mlsanity.com/paper/2502.16944/) (2025)
- [Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection](https://mlsanity.com/paper/2602.07892/) (2026)
- [EMA Policy Gradient: Taming Reinforcement Learning for LLMs with EMA Anchor and Top-k KL](https://mlsanity.com/paper/2602.04417/) (2026)
- [Learning to Control Summaries with Score Ranking](https://mlsanity.com/paper/2604.17197/) (2026)
- [Improving Interactive Diagnostic Ability of a Large Language Model Agent Through Clinical Experience Learning](https://mlsanity.com/paper/2503.16463/) (2025)
- [Generative Adversarial Post-Training Mitigates Reward Hacking in Live Human-AI Music Interaction](https://mlsanity.com/paper/2511.17879/) (2025)
- [General Exploratory Bonus for Optimistic Exploration in RLHF](https://mlsanity.com/paper/2510.03269/) (2025)
- [URL: Universal Referential Knowledge Linking via Task-instructed Representation Compression](https://mlsanity.com/paper/2404.16248/) (2024)
- [On the Transformations across Reward Model, Parameter Update, and In-Context Prompt](https://mlsanity.com/paper/2406.16377/) (2024)

## Abstract (arXiv)

Finetuning language models with reinforcement learning (RL), e.g. from human feedback (HF), is a prominent method for alignment. But optimizing against a reward model can improve on reward while degrading performance in other areas, a phenomenon known as reward hacking, alignment tax, or language drift. First, we argue that commonly-used test metrics are insufficient and instead measure how different algorithms tradeoff between reward and drift. The standard method modified the reward with a Kullback-Lieber (KL) penalty between the online and initial model. We propose Elastic Reset, a new algorithm that achieves higher reward with less drift without explicitly modifying the training objective. We periodically reset the online model to an exponentially moving average (EMA) of itself, then reset the EMA model to the initial model. Through the use of an EMA, our model recovers quickly after resets and achieves higher reward with less drift in the same number of steps. We demonstrate that fine-tuning language models with Elastic Reset leads to state-of-the-art performance on a small scale pivot-translation benchmark, outperforms all baselines in a medium-scale RLHF-like IMDB mock sentiment task and leads to a more performant and more aligned technical QA chatbot with LLaMA-7B. Code available at github.com/mnoukhov/elastic-reset.