Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
  • DailyAI Briefing: A Lean-checked Riemann proof, plus 4 extra
  • Anthropic Shares Drop 8.5% in Second Session Throughout AI Selloff – Unite.AI
AI-trends.todayAI-trends.today
Home»Tech»Hexo Labs Releases the Source Code for SIA, a Self-Improving agent that updates both the harness and model weights

Hexo Labs Releases the Source Code for SIA, a Self-Improving agent that updates both the harness and model weights

Tech By Gavin Wallace29/05/20267 Mins Read
Facebook Twitter LinkedIn Email
This AI Paper Introduces MMaDA: A Unified Multimodal Diffusion Model
This AI Paper Introduces MMaDA: A Unified Multimodal Diffusion Model
Share
Facebook Twitter LinkedIn Email

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.

https://arxiv.org/pdf/2605.27276

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

Hexo Labs · Open Source (MIT)

AI that improves itself

The latest updates on Harnesses and Weight

This loop edits the agent’s model weights and scaffold without requiring further human tweaking.

gpt-oss-120b
LoRA rank 32
3 benchmarks
Claude Sonnet 4.6 agents

The Gap

The two silos operate independently

Harness school

The scaffolding can be edited

The meta-agent can rewrite the prompts, tools and logic. Model weights remain fixed.

Training for the test

Weights can be edited

The model is updated by a RL pipeline based on the feedback. The harness remains fixed.


SIA fills in the gaps by moving both levers within one loop.

Anatomy

What SIA is

  • Harness (scaffolding) This includes the prompt for the system, logic of tool dispatch, policy on retries, and code to extract answers.
  • Weights: LoRA’s rank 32, with the parameters of your model.
  • Three LLM components Drive the loop: A Meta-Agent (a meta-agent), a task-specific agent, and a feedback-agent.

The Loop

One loop, two levers

Feedback Agent reads each trajectory after every run and selects a single action.

Action A

Harness Update

The scaffold can be rewritten. The weights will not move.

Action B

Weight Update

LoRA Weights can be used for training. Fix the scaffold.


Both levers are interlocking, but not sequentially.

You can also find out more about the Evidence

Benchmark results

Work The initials of the word “Initials” are: Prev. SOTA SIA-H SIA-W+H
LawBench – top-1 access 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

On all three tests, SIA W+H (harness and weights) won over SIA H (harness alone).

The Mechanism

What the Feedback-Agent does to choose its next move

  • LawBench: It used GAE with PPO to reward clean outcomes. Accuracy was 70.1%.
  • TriMul: It used the entropic weighting because most kernels do not compile. Runtime hit 1,017 µs.
  • Denoising: The value network was eliminated by using GRPO. The score rose from 0.189 to 0.289.
  • Also Available: The REINFORCE system includes KL to base, DPOs, and behavioural cloning based on the N-best.

RQ2

How each lever works

Harness

Externalised changes

Improvements in software engineering: New tools, tighter Parsers and retry logic.

Weighing

Internalised knowledge

No prompt is given for domain knowledge: The H100 kernel pattern, which rounds integers.


Weight updates alter what the model understands.

You Can Read The Truth

Keep in mind the following limitations

  • Two levers are used to optimize the Same fixed verifierRisking the Goodhart Effect.
  • The verifier can show a strong point but still be unable to detect it. The fragile Under Disturbance.
  • Newspaper reports Three TasksResults of algorithm-based selections are delayed.
  • The separate 350× superintelligence The newspaper does not mention the claim made in its coverage of launch.

Get Started

It’s up to you

Open source software under MIT license at hexo/sia. Built using gpt oss 120b and LoRA rank number 32.

# Install the Claude Backend
Pip Install 'sia-agent[claude]'
Export ANTHROPIC_API_KEY="..."

# Perform 5 self improvement generations for a single task
Sia -- Task Lawbench --max_gen 5 --run_id 1

In the box, you will find four bundles of tasks, including gpqa (lawbench), longcot-chess and spaceship-titanic.


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


ar code dat Update x
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026

OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min

21/11/2026

Meet Pondering Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac

21/11/2026

Baseten Releases Bison-1 and Groundhog-1, Topping Choice Mannequin Index

21/11/2026
Top News

Meta Superintelligence Lab’s Researchers are Already Departing

The Inside Story of the AI Summit where China presented its AI agenda to the world

There are 85 seconds left until midnight. Find Out What This Means

A dark-money campaign pays influencers to portray Chinese AI in a negative light

Does AI have the ability to diagnose Alzheimer’s by looking at your retina? Eric Topol Is Hopeful

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

Momfluencers Are Pitching AI as a Better ‘Coparent’ Than Men

08/06/2026

Photon releases Spectrum, an open-source TypeScript framework that deploys AI agents directly to iMessages, WhatsApp and Telegram

22/04/2026
Latest News

OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI

22/11/2026

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.