Faroq AL-Tam

Shuyin Ouyang, Zhaozhi Qian, Faroq AL-Tam et al.

Reinforcement Learning (RL) is an important paradigm for aligning Diffusion Language Models (DLMs) toward functional correctness in code generation. However, these models often encounter a ``capability cliff'' on complex tasks, where execution-based semantic rewards become too low to provide a viable learning signal. In this paper, we present a systematic empirical study of RL post-training for diffusion-based code generation along three axes: reward design, hint-conditioned sampling, and task difficulty. We investigate the effectiveness of execution-free rewards as alternatives to traditional unit-test execution, the role of training-time hint-conditioned diffusion sampling in mitigating exploration bottlenecks, and the impact of these design choices varies across tasks with different difficulty levels. Across HumanEval, MBPP, and LiveCodeBench, we find that static checking is the strongest overall standalone execution-free reward in our setting, especially improving DiffuCoder from 53.9 to 67.1 on HumanEval and from 14.9 to 15.5 on LiveCodeBench while reducing rollout time by 9.4\%. We further find that moderate AST-based hinting is most useful on harder benchmarks, while the best reward design depends strongly on task difficulty: similarity-based rewards are more effective on easier subsets, whereas static checking is more reliable on harder subsets where execution rewards are low. These findings suggest that reward design and training guidance substantially affect diffusion RL performance in our evaluated code-generation setting.

Ignacio Iacobacci, Zhaozhi Qian, Faroq AL-Tam et al.

Recently, a new wave of thinking-capable Large Language Models has emerged, demonstrating exceptional capabilities across a wide range of reasoning benchmarks. Early studies have begun to explore how the amount of compute in terms of the length of the reasoning process, the so-called thinking budget, impacts model performance. In this work, we propose a systematic investigation of the thinking budget as a key parameter, examining its interaction with various configurations such as self-consistency, reflection, and others. Our goal is to provide an informative, balanced comparison framework that considers both performance outcomes and computational cost. Among our findings, we discovered that simply increasing the thinking budget is not the most effective use of compute. More accurate responses can instead be achieved through alternative configurations, such as self-consistency and self-reflection.

Athul M. Mathew, Arshad Ali Khan, Thariq Khalid et al.

Gaze target detection (GTD) is the task of predicting where a person in an image is looking. This is a challenging task, as it requires the ability to understand the relationship between the person's head, body, and eyes, as well as the surrounding environment. In this paper, we propose a novel method for GTD that fuses multiple pieces of information extracted from an image. First, we project the 2D image into a 3D representation using monocular depth estimation. We then extract a depth-infused saliency module map, which highlights the most salient (\textit{attention-grabbing}) regions in image for the subject in consideration. We also extract face and depth modalities from the image, and finally fuse all the extracted modalities to identify the gaze target. We quantitatively evaluated our method, including the ablation analysis on three publicly available datasets, namely VideoAttentionTarget, GazeFollow and GOO-Real, and showed that it outperforms other state-of-the-art methods. This suggests that our method is a promising new approach for GTD.