Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning

Recent progress in large language models (LLMs) highlights the power of scaling test-time compute to achieve strong performance on complex tasks, such as mathematical reasoning and code generation. This raises a critical question: how should model training be modified to optimize performance under a subsequent test-time compute strategy and budget? To explore this, we focus on pass@N, a simple test-time strategy that searches for a correct answer in $N$ independent samples. We show, surprisingly, that training with cross-entropy (CE) loss can be ${\it misaligned}$ with pass@N in that pass@N accuracy ${\it decreases}$ with longer training. We explain the origins of this misalignment in terms of model overconfidence induced by CE, and experimentally verify our prediction of overconfidence as an impediment to scaling test-time compute via pass@N. Furthermore we suggest a principled, modified training loss that is better aligned to pass@N by limiting model confidence and rescuing pass@N test performance. Our algorithm demonstrates improved mathematical reasoning on MATH and MiniF2F benchmarks under several scenarios: (1) providing answers to math questions; and (2) proving theorems by searching over proof trees of varying shapes. Overall our work underscores the importance of co-designing two traditionally separate phases of LLM development: training-time protocols and test-time search and reasoning strategies.

Reinforcement Learning based Control of Imitative Policies for Near-Accident Driving

Autonomous driving has achieved significant progress in recent years, but autonomous cars are still unable to tackle high-risk situations where a potential accident is likely. In such near-accident scenarios, even a minor change in the vehicle's actions may result in drastically different consequences. To avoid unsafe actions in near-accident scenarios, we need to fully explore the environment. However, reinforcement learning (RL) and imitation learning (IL), two widely-used policy learning methods, cannot model rapid phase transitions and are not scalable to fully cover all the states. To address driving in near-accident scenarios, we propose a hierarchical reinforcement and imitation learning (H-ReIL) approach that consists of low-level policies learned by IL for discrete driving modes, and a high-level policy learned by RL that switches between different driving modes. Our approach exploits the advantages of both IL and RL by integrating them into a unified learning framework. Experimental results and user studies suggest our approach can achieve higher efficiency and safety compared to other methods. Analyses of the policies demonstrate our high-level policy appropriately switches between different low-level policies in near-accident driving situations.

Real-Time Panoptic Segmentation from Dense Detections

Panoptic segmentation is a complex full scene parsing task requiring simultaneous instance and semantic segmentation at high resolution. Current state-of-the-art approaches cannot run in real-time, and simplifying these architectures to improve efficiency severely degrades their accuracy. In this paper, we propose a new single-shot panoptic segmentation network that leverages dense detections and a global self-attention mechanism to operate in real-time with performance approaching the state of the art. We introduce a novel parameter-free mask construction method that substantially reduces computational complexity by efficiently reusing information from the object detection and semantic segmentation sub-tasks. The resulting network has a simple data flow that does not require feature map re-sampling or clustering post-processing, enabling significant hardware acceleration. Our experiments on the Cityscapes and COCO benchmarks show that our network works at 30 FPS on 1024x2048 resolution, trading a 3% relative performance degradation from the current state of the art for up to 440% faster inference.

Learning to Fuse Things and Stuff

We propose an end-to-end learning approach for panoptic segmentation, a novel task unifying instance (things) and semantic (stuff) segmentation. Our model, TASCNet, uses feature maps from a shared backbone network to predict in a single feed-forward pass both things and stuff segmentations. We explicitly constrain these two output distributions through a global things and stuff binary mask to enforce cross-task consistency. Our proposed unified network is competitive with the state of the art on several benchmarks for panoptic segmentation as well as on the individual semantic and instance segmentation tasks.