Jayanth Srinivasa

LLMs struggle to apply programming-language semantics, relying instead on learned priors, especially as program complexity rises.

Critique-aware training can boost LLM agent reliability by over 10% in complex, stateful environments, transforming how we approach tool-calling tasks.

Ground-truth access in the task-generating proposer can paradoxically *accelerate* self-play collapse, suggesting that ungrounded proposers might be more stable partners for self-consistency solvers.

Open-source LLM agents can get a 27% performance boost in tool use by strategically injecting context tailored to address their most common failure modes.

Despite advances in LLMs, human-AI collaboration still significantly outperforms AI-only agents in domain-specific data science tasks, proving that human expertise remains crucial.

Injecting demonstrations with a carefully annealed probability can drastically improve exploration in RLVR, even for tasks requiring novel reasoning or domain-specific knowledge.

LLMs struggle to consistently use tools in dynamic environments, but a simple input reformulation strategy can boost performance by over 16% compared to standard methods like ReAct.