Building self-improving AI systems

What it is A technical talk on the emerging frontier of self-improving AI systems, hosted by the Lossfunk Event Calendar and delivered by Amit Sharma, an AI researcher building a lab focused on cybersecurity and self-improving AI. The session is held in Bengaluru and streamed online via Google Meet.

Who it's for AI/ML researchers, applied scientists, engineers building agentic systems, and technical founders working on reasoning models, RL training pipelines, and agent reliability. The content is research-forward but grounded in executable frameworks, making it useful for practitioners shipping LLM-based products.

What to expect As large language models are applied to complex domains, the frontier is shifting from models trained once to systems that improve themselves. Amit decomposes self-improvement into three axes:

• Environment generation - introducing executable counterfactuals, a framework that operationalizes counterfactual reasoning as code, enabling scalable generation of RL environments with controllable difficulty and verifiable answers.
• RL training - training on these environments induces the core behaviors of counterfactual reasoning: abduction, intervention, and prediction. Unlike supervised fine-tuning, this generalizes out of domain to new code structures and math word problems.
• Harness evolution - presenting interwhen, which compiles natural-language policies into formal verifiers that certify each step of an agent's reasoning at runtime. This raises the pass^4 reliability of a 30B open-weight model on τ²-bench Telecom from 32% to 87%.

These axes compose into a single self-improvement loop, closing with a central research question: how to perform RL jointly with harness components and self-evolve both the model and its harness.

Pre-reads

• interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verification
• Executable Counterfactuals: Improving LLMs' Causal Reasoning Through Code

About the speaker Amit Sharma spent the past decade at Microsoft Research working on AI reasoning and causal inference. His work has produced foundational contributions to causal reasoning with applications for generalization, explainability, and reasoning in AI systems. He developed the DiCE algorithm for counterfactual explanation and refutation methods for evaluating causal estimates, widely adopted across academia and industry. He is also co-founder of PyWhy, an open-source ecosystem involving Carnegie Mellon University, Microsoft, Amazon, and others, advancing scalable causal ML tools.

Join online meet.google.com/onq-amgt-pgz