SIA: Self Improving AI with Harness & Weight UpdatesarXiv
A language-model agent that simultaneously modifies task-specific scaffolding and model weights. Achieves 25.1% over prior SOTA on LawBench and 12.4% faster GPU kernels.
Yogendra Manawat
AI researcher building Unintellect: small, specialized AI for real business workflows.
ICML 2026
4 papers accepted at ICML 2026, held in Seoul, South Korea - including SIA, our self-improving AI system.
At the ICML 2026 venue, Seoul, South Korea
View all photosCurrent Research Focus
I study how to make small language models (SLMs) capable enough for useful work on the hardware people already have. The goal is to make more AI tasks practical on hardware people and teams already have. That work informs what I'm building at Unintellect.
- Efficient inference on everyday hardware: Quantization, efficient decoding, and lower memory use to make capable models practical on phones and laptops.
- Making small models more capable: Distillation, better training data, and lightweight adaptation to help a small model do more with limited compute.
Why small models are a research problem
A smaller model uses less memory and compute, but has less capacity to handle difficult tasks. The research challenge is to improve what it can do within those limits, through better data, training, and inference, so useful work can run on everyday hardware.
Research
View all papers →- SIA-W: Self-Improving Agents with Test-Time Weight UpdatesICML 2026
Autonomous self-refinement via evolving agent structure and test-time reinforcement learning. +16pp on LawBench, -19% GPU kernel runtime.
- Adaptive Proxy Evaluation for Autonomously Improving ML AgentsICML 2026
Addresses the cost/reliability tradeoff in proxy evaluations. MLEvolve achieved SOTA MAE of 0.1354 on MLE-bench within 12 hours.
- Socrates: Structured Questioning Unlocks Latent Knowledge in AI Research AgentsICML 2026
A two-agent system pairing a Scientist with an advisor that can only ask questions. Improved Kaggle test scores on 4/5 MLE-bench tasks with a mean increase of ~56%.
- AIE-Bench: Benchmarking Agents That Build AgentsICML 2026
A benchmark for evaluating whether an AI agent can modify another agent to improve it, covering meta-improvement and self-improvement scenarios.
Experience
UnintellectAI Researcher
Building Unintellect to make AI useful in business workflows, with systems designed around the work they need to do and a focus on speed and affordability.
- AI built to fit real-world workflows
- Fast, affordable systems for practical use
Hexo LabsSenior Research Scientist
Leading AI research and high-impact programs end-to-end, from architecture and novel research to production delivery.
- 4 papers @ ICML 2026
- Lead author of SIA; co-built the system
- Led tech across multiple AI projects
- Owned end-to-end delivery for client programs
AI CallerAI / Backend Engineer
Built a real-time AI calling system before audio-to-audio models existed. Engineered low-latency voice pipelines from scratch and took it from zero to revenue.
- $2,000+ MRR within 2 months
- Pre audio-to-audio era
- Ultra-low latency voice AI
- Sole engineer on critical infrastructure
Who I've Worked With
Organizations I've worked with across AI research, engineering, and product development.
DPIIT
IP India
Bito
Soliton
Atomicwork
Dashtoon