Large Language Models (LLMs) and Vision-Language Models (VLMs) increasingly rely on compute-intensive post-training, including supervised fine-tuning (SFT) and reinforcement learning (RL), as well as inference-time scaling through long chains of thought (CoT), tool use, and agentic interaction. During post-training, certain frontier systems are reported to use compute at scales comparable to those of pre-training. Meanwhile, during deployment, agentic reasoning can consume hundreds of thousands to millions of tokens per task through iterative reasoning, tool use, and interaction. As a result, computational efficiency has become a central bottleneck in both post-training and inference.

The objective of this thesis proposal is to develop principled methods for improving the computational efficiency of LLM/VLM post-training and reasoning while maintaining strong performance, by reducing data requirements, training computation, and inference-time computation. Toward this objective, the proposal investigates three complementary dimensions of efficiency: (i) data efficiency, (ii) training efficiency, and (iii) inference efficiency.

First, we introduce a token-selective hierarchical data selection method for SFT. The method identifies high-quality training examples using token-level informativeness and local robustness. By selecting only a small fraction of a large SFT dataset while achieving comparable or even better downstream performance, it improves both data and training efficiency.

Second, we introduce an agentic data engine for synthesizing SFT data that teaches VLM agents to logically think with images. VLMs trained on the synthesized data develop substantially stronger visual search capabilities after SFT and RL while requiring less training data, and they achieve more accurate and efficient tool use. This approach therefore improves data, training, and inference efficiency simultaneously.

Third, we plan to develop a plug-and-play credit assignment method for RL of LLMs. The method will leverage mechanistic interpretability to estimate credit over intermediate reasoning steps, avoiding expensive partial-trajectory labeling based on Monte Carlo sampling. By providing more efficient supervision for long-horizon reasoning, the approach aims to improve both data and training efficiency.

Finally, we plan to develop a training-free, mechanistic-interpretability-inspired method for steering LLM reasoning at inference time. The goal is to reduce redundant and repetitive reasoning while preserving task performance, thereby improving inference efficiency without additional training.

Together, these directions form a unified effort to improve the efficiency of LLM/VLM post-training and reasoning across data, training, and inference, providing complementary approaches to reducing computational overhead throughout the model development and deployment pipeline.