We introduce DINGO, a novel constrained inference method for diffusion-based large language models that enables efficient generation while satisfying user-defined constraints. Our approach provides a principled way to incorporate constraints into the diffusion process, enabling more controllable and reliable LLM generation.

We show theoretically that constraining LLM generation to a fixed output grammar can reduce reasoning capabilities and by augmenting the grammar with additional production rules for CoT reasoning steps, we can preserve LLM expressivity. Building on these theoretical results, we introduce CRANE, a constrained decoding algorithm that only enforces the grammar when generating final answers and intermediate expressions, while keeping reasoning steps unconstrained. CRANE boosts accuracy by up to 10 percentage points on first-order logic generation and other symbolic reasoning tasks.

IterGen is a decoding algorithm which leverages grammar-based backtracking and selective rejection sampling to efficiently enforce user-defined semantic constraints into LLM output. IterGen improves LLM-generated SQL accuracy by 18% and eliminates LLM privacy leakage.

SynCode is a novel framework for the grammar-guided generation of Large Language Models (LLMs) that is scalable to general-purpose programming languages and has soundness and completeness guarantees. SynCode reduces syntax errors by 96-100% for various languages (JSON, Python, Go) and enables 1.5x-10x faster LLM inference than existing approaches.

We propose PET, a novel pessimistic reward fine-tuning method, to learn a pessimistic reward model robust against reward hacking in offline reinforcement learning from human feedback (RLHF). Our method shows that when optimizing a policy on a pessimistic reward model fine-tuned through PET, reward hacking can be prevented without relying on any regularization, enabling agents to greedily search for high-reward policies without suffering from reward hacking.

To address challenges from the risk of reward hacking and the complexity of reinforcement learning during preference-based reinforcement learning, we develop a novel two-step learning method called PRC. The high-level idea is to limit the reinforcement learning agent to optimize over a constrained action space that excludes out-of-distribution state-actions, which are unreliable and increase the complexity of the reinforcement learning problem.

We propose CodeARC, the Code Abstraction and Reasoning Challenge, a new evaluation framework where agents interact with a hidden target function by querying it with new inputs, synthesizing candidate functions, and iteratively refining their solutions using a differential testing oracle. We construct the first large-scale benchmark for general-purpose inductive program synthesis, featuring 1114 functions. Among 18 models evaluated, o3-mini performs best with a success rate of 52.7%, highlighting the difficulty of this task.

We introduce CoRNStack, a large-scale, high-quality contrastive training dataset for code that spans multiple programming languages. We demonstrate that contrastive training of embedding models and LLM re-rankers using CoRNStack leads to state-of-the-art performance across a variety of code retrieval tasks. Notably, our lightweight retriever + re-ranker achieves state-of-the-art repoistory level bug localization on SWE-Bench over top automated software development frameworks.

We present SweRank, a novel approach for software issue localization that leverages code ranking techniques to identify relevant code segments for bug fixing. Our method significantly improves localization accuracy by effectively ranking code files and functions based on their relevance to reported software issues.

We present a reinforcement learning framework that trains LLMs using Proximal Policy Optimization (PPO) to optimize assembly code performance. Our model, Qwen2.5-Coder-7B-PPO, achieves 96.0% test pass rates and an average speedup of 1.47x over the gcc -O3 baseline on a benchmark of 8,072 real-world programs, demonstrating that reinforcement learning can unlock the potential of LLMs to serve as effective optimizers for assembly code performance.

We develop a method, called TAR, for building tamper-resistant safeguards into open-weight LLMs such that adversaries cannot remove the safeguards even after thousands of steps of fine-tuning. In extensive evaluations and red teaming analyses, we find that our method greatly improves tamper-resistance while preserving benign capabilities.

We introduce a GPU-accelerated, Branch-and-Bound-based verifier RABBit for formally verifying hyperproperties, such as ensembles, conformal prediction, and robustness, defined over multiple executions of Deep Neural Networks. RABBit improves neural network verified accuracy by 8% over state-of-the-art verifiers within the same compute budget.

We present IRS, the first probabilistic approach for 5x faster robustness re-certification of Deep Neural Networks after model compression (pruning, quantization) or fine-tuning

We present the first study of the robustness of existing watermarking techniques on code generated by large language models and propose a parsing-based algorithm that easily removes these watermarks via semantic preserving transformations of the code.

We propose efficient deterministic formal verifiers to speed up DNN re-verification after pruning, quantization, or fine-tuning.