About

Process Level LLM Reasoning and Evaluation

Reasoning quality cannot always be captured by the final answer. A correct result may contain weak intermediate reasoning, while fixing one incorrect step does not necessarily recover the rest of the solution. My work examines where reasoning errors begin, how they affect later steps, and what happens after an intermediate failure is corrected.

This motivates process level evaluation that separates the quality of individual steps from the consistency of the full reasoning process and the final outcome. It also raises questions about verification, error propagation, and recovery after a failure has already influenced later steps. The goal is to understand what a model can actually do rather than treating answer accuracy as a complete measure of reasoning ability.

Mathematical Reasoning and Problem Solving

Mathematical problem solving depends on several capabilities that are often compressed into one accuracy number. Problem understanding, formulation, planning, action selection, computation, verification, revision, and recovery can each succeed or fail in different ways.

I want to understand how these components interact and which capabilities a model genuinely possesses. Computation is one current focus, especially when a capability that works in a simple setting becomes unreliable inside a longer reasoning context.

The same framework can later extend to planning, action selection, formulation, verification, and other parts of the reasoning process. The larger question is how these abilities combine into a stable problem solving system rather than appearing only as isolated strengths.

Reliable Data and Model Evaluation

Reliable evaluation depends on more than the model being tested. Data quality, labeling decisions, supporting evidence, human review, and metric design can all change the conclusions we draw about model capability. I am interested in evaluation that preserves enough information to inspect why a prediction was considered correct, where disagreements come from, and which failure patterns are hidden by aggregate scores. My goal is to make these factors easier to inspect and audit, so that evaluation reflects what a model can actually do rather than what a benchmark happens to reward.

Grounded Vision Language and Multimodal Understanding

My interest in vision comes from both research and production systems. I care about how models recognize objects and actions, understand temporal structure in video, and connect visual evidence with language, context, and downstream decisions. In multimodal settings, a strong answer is not enough if the model cannot be traced back to the image regions, events, or relationships that support that answer.

I am interested in separating failures in perception, temporal understanding, grounding, and higher level reasoning. This includes questions around video understanding, vision language models, visual grounding, multimodal evaluation, and how language changes the interpretation of visual evidence. I also care about the practical side of these systems, including data quality, model adaptation, efficient inference, and deployment constraints when visual models must operate reliably outside a benchmark.

Test Time Reasoning and Adaptive Computation

Additional computation is useful only when it changes the quality of the reasoning process. More tokens or more internal computation do not automatically mean that a model is reasoning better. I am interested in when a model should continue, verify an intermediate result, revise an earlier decision, allocate more internal computation, or stop. The central question is whether extra computation produces a meaningful change in capability rather than only increasing cost.