S Akash
22 ยท research at latentforce
i work across research and engineering at latentforce, where we’re building latentstack, a software factory for people and coding agents to ship together. if you're working on similar problems, i'd enjoy comparing notes.
i studied electrical engineering at IIT Patna and came to reinforcement learning through bandits: problems where an agent has to decide well while the information it needs is missing.
my bachelor’s thesis was on multi-dueling bandits, built around a simple reduction from multi-way winner feedback to ordinary duels. the learner only sees which option won, and still has to work out which one is best.
at latentforce the setting changed, but the problem did not. a coding agent rarely has everything it needs: the code is too large to read in full, and the reasons behind it live in team chats, meetings, and past sessions.
the question underneath all of it: out of the code, the conversations, and everything a team has decided, what does an agent actually need to know?
research
best-of-both-worlds multi-dueling bandits RLC ’26
a learner picks several options each round and sees only the winner. our algorithms do well whether preferences stay stable or turn adversarial.
fairness in two-player zero-sum games with bandit feedback preprint
a zero-sum game where every action must keep a minimum chance of being played. we bound what this fairness costs and how well it can be learned.