Chao Wang

Shaojie Wang, Jinghui Wang, Yinghan Cui et al.

In agentic LLM scenarios, an agent's interaction process during a single rollout often exhibits branching behaviors. Due to memory retrieval and concurrent tool executions at certain decision points, the token trajectory of one task evolves into a tree-like structure rather than a linear sequence. However, current training pipelines decompose such tree-structured trajectories into separate linear segments, treating each branch as an independent sequence. As a result, shared prefixes across these branches are repeatedly recomputed during both forward and backward passes. To address this inefficiency, we propose Tree Training, a paradigm that computes each shared prefix only once and reuses its intermediate results across related branches during both forward and backward passes, substantially improving computation efficiency in large-scale agentic training. This is achieved via (i) Tree Packing, which efficiently reuses shared computations across trajectories, and (ii) Gradient Restoration, which ensures correct gradient propagation across reused prefixes. Experiments on multiple open-source models demonstrate up to 3.9x reduction in total training time, enabling more efficient agentic LLM SFT and RL training.

Jingbo Wang, Chungha Sung, Mukund Raghothaman et al.

We propose a data-driven method for synthesizing a static analyzer to detect side-channel information leaks in cryptographic software. Compared to the conventional way of manually crafting such a static analyzer, which can be labor intensive, error prone and suboptimal, our learning-based technique is not only automated but also provably sound. Our analyzer consists of a set of type-inference rules learned from the training data, i.e., example code snippets annotated with ground truth. Internally, we use syntax-guided synthesis (SyGuS) to generate new features and decision tree learning (DTL) to generate type-inference rules based on these features. We guarantee soundness by formally proving each learned rule via a technique called Datalog query containment checking. We have implemented our technique in the LLVM compiler and used it to detect power side channels in C programs. Our results show that, in addition to being automated and provably sound during synthesis, the learned analyzer also has the same empirical accuracy as two state-of-the-art, manually crafted analyzers while being 300X and 900X faster, respectively.