# SakanaAI/self-adaptive-llms

A Self-adaptation Framework🐙 that adapts LLMs for unseen tasks in real-time!

Repository: https://github.com/SakanaAI/self-adaptive-llms
Canonical: https://ross.abutalabs.com/products/self-adaptive-llms
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-01-30T03:48:47+00:00

## Health v2 (maintenance only)
Score: 23/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 4, release rhythm 35, longevity 45
- inputs: {"age_days": 639, "days_push": 580, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1225, forks 142 (observed 2026-08-28T04:04:03.030619+00:00)

## What it is
Transformer² is a research framework from SakanaAI that adapts large language models to unseen tasks in real-time by selectively adjusting singular components of weight matrices. It uses a two-pass inference mechanism with a task dispatch system and reinforcement-learning-trained expert vectors that are dynamically mixed per prompt.

## Use cases
- adapt an LLM to unseen tasks in real-time without full fine-tuning
- train task-specific expert vectors with reinforcement learning
- evaluate self-adaptive LLM behavior on benchmarks
- research lightweight parameter-efficient LLM adaptation
- dynamically mix expert vectors based on incoming prompt task type

## When to choose
- you are doing research on self-adaptive or parameter-efficient LLM adaptation
- you want to avoid computationally expensive static fine-tuning for diverse tasks
- you want to reproduce the Transformer² paper results

## When to avoid
- you need a production-ready, well-supported LLM serving framework
- you need simple standard fine-tuning rather than adaptive expert-vector mixing
- you lack GPU resources or are not working in a research context

## Facets
- artifact type: library
- maturity: experimental
- function: machine-learning, llm-training, reinforcement-learning, llm-inference
- domain: large-language-models, machine-learning, artificial-intelligence
- platform: python
- tags: self-adaptation, research, transformer-squared, fine-tuning, expert-vectors, svd, linux, gpu

## Member repositories
- SakanaAI/self-adaptive-llms (main) score 23

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:03.030619+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:15:23.049810+00:00, confidence not recorded.
  - readme: https://github.com/SakanaAI/self-adaptive-llms (fetched 2026-08-28T04:04:03.030619+00:00, sha 09f527b7b3d1)
- Data as of 2026-08-30T08:39:29.467469+00:00.