Pith. sign in

REVIEW 10 cited by

ReFT: Representation Finetuning for Language Models

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2404.03592 v3 pith:RGIHND7E submitted 2024-04-04 cs.CL cs.AIcs.LG

Zhengxuan Wu , Aryaman Arora , Zheng Wang , Atticus Geiger , Dan Jurafsky , Christopher D. Manning , Christopher Potts This is my paper
classification cs.CLcs.AIcs.LG
keywords reftfinetuningmethodsrepresentationsefficiencyfamilyinterventionslearn
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Parameter-efficient finetuning (PEFT) methods seek to adapt large neural models via updates to a small number of weights. However, much prior interpretability work has shown that representations encode rich semantic information, suggesting that editing representations might be a more powerful alternative. We pursue this hypothesis by developing a family of Representation Finetuning (ReFT) methods. ReFT methods operate on a frozen base model and learn task-specific interventions on hidden representations. We define a strong instance of the ReFT family, Low-rank Linear Subspace ReFT (LoReFT), and we identify an ablation of this method that trades some performance for increased efficiency. Both are drop-in replacements for existing PEFTs and learn interventions that are 15x--65x more parameter-efficient than LoRA. We showcase LoReFT on eight commonsense reasoning tasks, four arithmetic reasoning tasks, instruction-tuning, and GLUE. In all these evaluations, our ReFTs deliver the best balance of efficiency and performance, and almost always outperform state-of-the-art PEFTs. We release a generic ReFT training library publicly at https://github.com/stanfordnlp/pyreft.

Discussion (0). Sign in to comment.

Forward citations

Cited by 10 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Agent Security Needs Redefinition through a Holistic Framework

    cs.CR 2026-07 conditional novelty 6.0 of 10

    Agent security should be redefined around four contextual authorization properties instead of the content of the action performed.

  2. You Had One Job: Per-Task Quantization Using LLMs' Hidden Representations

    cs.CL 2025-11 reject novelty 6.0 of 10

    TAQ estimates per-layer importance from hidden representations and output sensitivity on task calibration data to allocate mixed precision in a training-free PTQ setting, outperforming task-agnostic baselines on accur...

  3. Infogen: Generating Complex Statistical Infographics from Documents

    cs.CL 2025-07 conditional novelty 6.0 of 10

    Infogen generates complex statistical infographics from text via a metadata-then-code pipeline, and the new Infodat benchmark shows it outperforming GPT-4o and fine-tuned open LLMs on the authors' metrics.

  4. Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuning

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Fine-tuning a single transformer layer with activation targets scaled by each query's projection onto a refusal direction reduces over-refusal on four benchmarks while preserving safety and general capability.

  5. Improved Representation Steering for Language Models

    cs.CL 2025-05 conditional novelty 6.0 of 10

    RePS, a reference-free bidirectional preference optimization objective, improves representation steering and suppression for Gemma models, outperforming language-modeling objectives and approaching prompting performance.

  6. First-Order Predictable but Pairwise Fragile: Local Task Adaptation in Trained Transformers

    cs.LG 2026-07 conditional novelty 5.0 of 10

    Single task-vector perturbations around a multitask LoRA point are first-order linear to 1e-2, but pairwise update-order sensitivity is set by a per-pair Lie bracket and can appear inside that window.

  7. Dynamically Scaled Activation Steering

    cs.LG 2025-12 conditional novelty 5.0 of 10

    DSAS uses a per-token logistic gate to scale any activation-steering method, improving the toxicity-versus-quality Pareto front in LLMs and diffusion models.

  8. Enhancing Chain-of-Thought Reasoning with Critical Representation Fine-tuning

    cs.CL 2025-07 conditional novelty 5.0 of 10

    CRFT selects critical internal representations via attention and saliency scores and fine-tunes only them, improving GSM8K accuracy over ReFT from 29.0% to 32.8% on LLaMA-2-7B.

  9. Drop Dropout on Single-Epoch Language Model Pretraining

    cs.CL 2025-05 conditional novelty 5.0 of 10

    Removing dropout from single-epoch language model pretraining improves downstream task performance and model editability across multiple architectures and benchmarks.

  10. Mitigating Catastrophic Forgetting in Large Language Models with Forgetting-aware Pruning

    cs.LG 2025-09 conditional novelty 4.0 of 10

    FAPM prunes task vectors using a combined absolute and relative change magnitude criterion, reducing forgetting on general benchmarks to roughly a third of a percent while keeping downstream accuracy near the fine-tun...

Pith tools