Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Self-Adapting Language Models
Authors: Adam Zweiger, Jyo Pari, Han Guo, Yoon Kim, Pulkit Agrawal
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on knowledge incorporation and fewshot generalization show that SEAL is a promising step toward language models capable of self-directed adaptation in response to new data. Our results show that, following reinforcement learning (RL) training, finetuning on self-generated synthetic data improves question-answering performance on the no-passage-in-context variant of SQu AD [13] from 33.5% to 47.0%. Notably, self-generated data from SEAL outperforms synthetic data generated by GPT-4.1. We further evaluate SEAL on few-shot learning on a simplified subset of the ARC-AGI benchmark [14], where the model leverages a set of tools to autonomously select both synthetic data augmentations and optimization hyperparameters (e.g., learning rate, training epochs, selective loss computation over token types). |
| Researcher Affiliation | Academia | Adam Zweiger Jyothish Pari Han Guo Yoon Kim Pulkit Agrawal Massachusetts Institute of Technology EMAIL |
| Pseudocode | Yes | Algorithm 1 Self-Adapting LLMs (SEAL): Self-Edit Reinforcement Learning Loop 1: Input: LMθ, dataset D = {(C, τ)} 2: for outer iteration t = 1, 2, . . . do 3: Sample (C, τ) D 4: Generate self-edit SE LMθ( | C) 5: Inner Loop Update: θ t SFT(θt, SE) 6: Evaluate: Ans LMθ t( | τ) 7: Compute reward: r r(Ans, τ) 8: Update: θt+1 RL_Update(θt, r, SE) 9: end for |
| Open Source Code | Yes | Our website and code is available at https://jyopari.github.io/posts/seal. |
| Open Datasets | Yes | Our results show that, following reinforcement learning (RL) training, finetuning on self-generated synthetic data improves question-answering performance on the no-passage-in-context variant of SQu AD [13] from 33.5% to 47.0. |
| Dataset Splits | Yes | For the few-shot learning experiments... We curated a small set of ARC problems from the training and evaluation splits that are solvable with optimal TTT hyperparameters. Training Set: We selected 11 ARC tasks from the training set as the environment for RL optimization. Evaluation Set: We selected 8 distinct ARC problems from the evaluation set for measuring generalization performance. |
| Hardware Specification | Yes | We performed all training runs on a single A100, H100, or H200. |
| Software Dependencies | No | We use Deep Speed Ze RO-3 [83] for SFT in Re STEM training. We use v LLM [84] for efficient inference. |
| Experiment Setup | Yes | SFT for TTT was done with the following hyperparameters: Parameter Value Lo RA rank 128 Lo RA alpha 16 Learning rate N/A (model chooses it) num_train_epochs N/A (model chooses it) Lo RA adapters were inserted into the following modules: q_proj, v_proj, gate_proj, down_proj, up_proj For the RL SFT we used the following hyperparameters: Parameter Value lora_rank 16 lora_alpha 16 num_train_epochs 8 per_device_train_batch_size 5 learning_rate 5e-5 |