The majority of AI agents cease to improve once humans stop tuning them. Model is set. The scaffolding around the model is also fixed. Hexo Labs would like to do both moves at the same time. The release was released. SIA (Self-Improving AI) This week, MIT released a framework that is open source.
The research claims are narrow and concrete. SIA updates both the scaffold of the agent and the weights for the model in one loop.
What is SIA(Self-Improving Artificial Intelligence)?
SIA separates a job-specific agent into 2 parts. The harness is also known as the scaffold. This includes the prompt for system, logic of tool dispatch, policy on retries, and code to extract answers. Model weights are the second part.
The loop is driven by three LLM components. The Meta-Agent creates the first scaffold from task specifications and reference codes. The Task-Specific agent runs the task, logging every step. The Feedback Agent then reviews the full path and determines how to make changes.
The key concept is to make a decision. The Feedback-Agent selects one action from two options after each run. It can change the scaffold, while keeping weights fixed. Or, it can update the weights of the scaffold while remaining fixed.
Model base is openai/gpt.oss.120b. LoRA (a low-rank Adapter at 32) is used for weight updates. Both the Meta-Agent (also known as Feedback-Agent) and Claude Sonnet 4.6 run both on Claude Sonnet 4.6. Modal is the team’s RL system, which runs training on GPUs H100.
The team has two operational points, SIAH and SIAW+H. SIA only updates harnesses. Weight updates are added to SIA-W+H.
Benchmark Case
Three deliberately different domains were chosen to test SIA. All three domains showed the same pattern. Weigh updates improved the results beyond scaffold editing. “Initial” Before any feedback, the first step is to build the Meta-Agent model.
| Work | If you are unsure, please contact us. | Prev. SOTA | SIA-H (harness only) | SIA-W+H (harness + weights) |
|---|---|---|---|---|
| LawBench (top-1 account) | 13.5% | 45.0% | 50.0% | 70.1% |
| AlphaEvolve TriMul – Reward | 0.105 | 1.292 | 0.120 | 1.475 |
| Denoising | 0.048 | 0.240 | 0.241 | 0.289 |
LawBench’s task is the 191-class Chinese criminal offense classification. The Harness Iteration created a TFIDF pipeline with LinearSVC and reached a plateau of 50.0%. PPO weight updates then improved accuracy to 70%. It is a gain of 20,1 percentage points over the best harness only.
TriMul tasks require a custom CUDA Kernel on a GPU H100. This kernel is used to compute a key operation within AlphaFold2’s Evoformer Module. Scaffold edits reached a 1.14× speedup over baseline. The runtime was reduced from 12,483 microseconds to just 1,017 after weight updates. This is a reduction of 91.9% from the peak harness-only.
The same chart contains a warning. The coding agent Claude Code reached 1.50× on TriMul unaided, beating SIA-H’s 1.14×. SIA-W+H still led overall at 14.02×.
The agent uses MAGIC to tune the method of single-cell RNA imputing. It settles at 0.241 for its hyperparameters. A two-line increment was added to the first checkpoint of weight updates, which no scaffold could produce. This rounded the numbers to positive integers and raised the score from 0.189 to 0.289.
The Feedback Agent Picks its Move
SIA runs no fixed RL recipes. The Feedback Agent selects an algorithm for training based on what it sees.
LawBench’s reward is a pure outcome-based matrix, therefore it uses PPO and GAE. TriMul’s kernels were unable to be compiled, and so the entropic benefit weighting was used. The method is used to weight rare rollouts with high rewards. Denoising was done using GRPO which removes the entire value network.
Also, the research team lists REINFORCE using KL-to base, DPO or best-ofN behaviour cloning. Each map to a distinct reward shape and risk of failure.
Strengths What to Watch
Strengths:
- The authors have compared the two systems and found that this first method allows both scaffolding and weights to be edited in one loop.
- Gains consistent over previous SOTA in three domains unrelated to each other.
- Installable sia-agent under MIT and open source, bundled with four tasks.
- It is not necessary to have a schedule. Instead, the algorithm choice depends on rewards observed.
What to watch
- This research examines three specific tasks. Results on algorithm selection are not included.
- The Goodhart effect is coupled when both levers are used to optimize the same verifier.
- Researchers warn that the joint fixation point could be unstable under disturbance.
Marktechpost’s Visual Explainer
01 / 09
What you need to know
- SIA, the self-improving algorithm that improves an agent’s model and scaffold weights.
- The Feedback Agent reads the full run trajectory and then selects either a weight or harness update.
- Combining the two levers is superior to scaffold-only for all three tasks, including LawBench kernels, TriMul, and scRNA sequencing denoising.
- The weight update reveals domain knowledge not reached by the prompt.
- Open source (hexo/sia) under MIT, built on gpt oss-120b and LoRA rank 33.
Check out the Repo You can also find out more about the following: Research Paper. Also, feel free to follow us on Twitter Join our Facebook group! 150k+ ML SubReddit Subscribe Now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us

