Abstract:Audio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resource-constrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that Llama-SMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies confirm its effectiveness in expert activation, scalability, and noise robustness.
Abstract:Diffusion models have demonstrated strong generative capabilities across domains ranging from image synthesis to complex reasoning tasks. However, most inference-time scaling methods rely on fixed denoising schedules, limiting their ability to allocate computation based on instance difficulty or task-specific demands adaptively. We introduce the challenge of adaptive inference-time scaling-dynamically adjusting computational effort during inference-and propose Adaptive Bi-directional Cyclic Diffusion (ABCD), a flexible, search-based inference framework. ABCD refines outputs through bi-directional diffusion cycles while adaptively controlling exploration depth and termination. It comprises three components: Cyclic Diffusion Search, Automatic Exploration-Exploitation Balancing, and Adaptive Thinking Time. Experiments show that ABCD improves performance across diverse tasks while maintaining computational efficiency.
Abstract:Reinforcement learning (RL) has proven effective for fine-tuning large language models (LLMs), significantly enhancing their reasoning abilities in domains such as mathematics and code generation. A crucial factor influencing RL fine-tuning success is the training curriculum: the order in which training problems are presented. While random curricula serve as common baselines, they remain suboptimal; manually designed curricula often rely heavily on heuristics, and online filtering methods can be computationally prohibitive. To address these limitations, we propose Self-Evolving Curriculum (SEC), an automatic curriculum learning method that learns a curriculum policy concurrently with the RL fine-tuning process. Our approach formulates curriculum selection as a non-stationary Multi-Armed Bandit problem, treating each problem category (e.g., difficulty level or problem type) as an individual arm. We leverage the absolute advantage from policy gradient methods as a proxy measure for immediate learning gain. At each training step, the curriculum policy selects categories to maximize this reward signal and is updated using the TD(0) method. Across three distinct reasoning domains: planning, inductive reasoning, and mathematics, our experiments demonstrate that SEC significantly improves models' reasoning capabilities, enabling better generalization to harder, out-of-distribution test problems. Additionally, our approach achieves better skill balance when fine-tuning simultaneously on multiple reasoning domains. These findings highlight SEC as a promising strategy for RL fine-tuning of LLMs.
Abstract:Chain-of-Thought (CoT) reasoning has advanced the capabilities and transparency of language models (LMs); however, reasoning chains can contain inaccurate statements that reduce performance and trustworthiness. To address this, we introduce a new self-correction framework that augments each reasoning step in a CoT with a latent variable indicating its veracity, enabling modeling of all possible truth assignments rather than assuming correctness throughout. To efficiently explore this expanded space, we introduce Search Corrector, a discrete search algorithm over boolean-valued veracity assignments. It efficiently performs otherwise intractable inference in the posterior distribution over veracity assignments by leveraging the LM's joint likelihood over veracity and the final answer as a proxy reward. This efficient inference-time correction method facilitates supervised fine-tuning of an Amortized Corrector by providing pseudo-labels for veracity. The Amortized Corrector generalizes self-correction, enabling accurate zero-shot veracity inference in novel contexts. Empirical results demonstrate that Search Corrector reliably identifies errors in logical (ProntoQA) and mathematical reasoning (GSM8K) benchmarks. The Amortized Corrector achieves comparable zero-shot accuracy and improves final answer accuracy by up to 25%.
Abstract:In the context of continual learning, acquiring new knowledge while maintaining previous knowledge presents a significant challenge. Existing methods often use experience replay techniques that store a small portion of previous task data for training. In experience replay approaches, data augmentation has emerged as a promising strategy to further improve the model performance by mixing limited previous task data with sufficient current task data. However, we theoretically and empirically analyze that training with mixed samples from random sample pairs may harm the knowledge of previous tasks and cause greater catastrophic forgetting. We then propose GradMix, a robust data augmentation method specifically designed for mitigating catastrophic forgetting in class-incremental learning. GradMix performs gradient-based selective mixup using a class-based criterion that mixes only samples from helpful class pairs and not from detrimental class pairs for reducing catastrophic forgetting. Our experiments on various real datasets show that GradMix outperforms data augmentation baselines in accuracy by minimizing the forgetting of previous knowledge.
Abstract:We propose PhysicsFC, a method for controlling physically simulated football player characters to perform a variety of football skills--such as dribbling, trapping, moving, and kicking--based on user input, while seamlessly transitioning between these skills. Our skill-specific policies, which generate latent variables for each football skill, are trained using an existing physics-based motion embedding model that serves as a foundation for reproducing football motions. Key features include a tailored reward design for the Dribble policy, a two-phase reward structure combined with projectile dynamics-based initialization for the Trap policy, and a Data-Embedded Goal-Conditioned Latent Guidance (DEGCL) method for the Move policy. Using the trained skill policies, the proposed football player finite state machine (PhysicsFC FSM) allows users to interactively control the character. To ensure smooth and agile transitions between skill policies, as defined in the FSM, we introduce the Skill Transition-Based Initialization (STI), which is applied during the training of each skill policy. We develop several interactive scenarios to showcase PhysicsFC's effectiveness, including competitive trapping and dribbling, give-and-go plays, and 11v11 football games, where multiple PhysicsFC agents produce natural and controllable physics-based football player behaviors. Quantitative evaluations further validate the performance of individual skill policies and the transitions between them, using the presented metrics and experimental designs.
Abstract:This paper investigates the knowledge of language models from the perspective of Bayesian epistemology. We explore how language models adjust their confidence and responses when presented with evidence with varying levels of informativeness and reliability. To study these properties, we create a dataset with various types of evidence and analyze language models' responses and confidence using verbalized confidence, token probability, and sampling. We observed that language models do not consistently follow Bayesian epistemology: language models follow the Bayesian confirmation assumption well with true evidence but fail to adhere to other Bayesian assumptions when encountering different evidence types. Also, we demonstrated that language models can exhibit high confidence when given strong evidence, but this does not always guarantee high accuracy. Our analysis also reveals that language models are biased toward golden evidence and show varying performance depending on the degree of irrelevance, helping explain why they deviate from Bayesian assumptions.
Abstract:We study model confidence calibration in class-incremental learning, where models learn from sequential tasks with different class sets. While existing works primarily focus on accuracy, maintaining calibrated confidence has been largely overlooked. Unfortunately, most post-hoc calibration techniques are not designed to work with the limited memories of old-task data typical in class-incremental learning, as retaining a sufficient validation set would be impractical. Thus, we propose T-CIL, a novel temperature scaling approach for class-incremental learning without a validation set for old tasks, that leverages adversarially perturbed exemplars from memory. Directly using exemplars is inadequate for temperature optimization, since they are already used for training. The key idea of T-CIL is to perturb exemplars more strongly for old tasks than for the new task by adjusting the perturbation direction based on feature distance, with the single magnitude determined using the new-task validation set. This strategy makes the perturbation magnitude computed from the new task also applicable to old tasks, leveraging the tendency that the accuracy of old tasks is lower than that of the new task. We empirically show that T-CIL significantly outperforms various baselines in terms of calibration on real datasets and can be integrated with existing class-incremental learning techniques with minimal impact on accuracy.
Abstract:Modern Large Language Model serving system batches multiple requests to achieve high throughput, while batching attention operations is challenging, rendering memory bandwidth a critical bottleneck. The community relies on high-end GPUs with multiple high-bandwidth memory channels. Unfortunately, HBM's high bandwidth often comes at the expense of limited memory capacity, which reduces core utilization and increases costs. Recent advancements enabling longer contexts for LLMs have substantially increased the key-value cache size, further intensifying the pressures on memory capacity. The literature has explored KV cache quantization techniques, which commonly use low bitwidth for most values, selectively using higher bitwidth for outlier values. While this approach helps achieve high accuracy and low bitwidth simultaneously, it comes with the limitation that cost for online outlier detection is excessively high, negating the advantages. We propose Oaken, an acceleration solution that achieves high accuracy and high performance simultaneously through co-designing algorithm and hardware. To effectively find a sweet spot in the accuracy-performance trade-off space of KV cache quantization, Oaken employs an online-offline hybrid approach, setting outlier thresholds offline, which are then used to determine the quantization scale online. To translate the proposed algorithmic technique into tangible performance gains, Oaken also comes with custom quantization engines and memory management units that can be integrated with any LLM accelerators. We built an Oaken accelerator on top of an LLM accelerator, LPU, and conducted a comprehensive evaluation. Our experiments show that for a batch size of 256, Oaken achieves up to 1.58x throughput improvement over NVIDIA A100 GPU, incurring a minimal accuracy loss of only 0.54\% on average, compared to state-of-the-art KV cache quantization techniques.
Abstract:Reinforcement learning (RL) is a critical component of large language model (LLM) post-training. However, existing on-policy algorithms used for post-training are inherently incompatible with the use of experience replay buffers, which can be populated scalably by distributed off-policy actors to enhance exploration as compute increases. We propose efficiently obtaining this benefit of replay buffers via Trajectory Balance with Asynchrony (TBA), a massively scalable LLM RL system. In contrast to existing approaches, TBA uses a larger fraction of compute on search, constantly generating off-policy data for a central replay buffer. A training node simultaneously samples data from this buffer based on reward or recency to update the policy using Trajectory Balance (TB), a diversity-seeking RL objective introduced for GFlowNets. TBA offers three key advantages: (1) decoupled training and search, speeding up training wall-clock time by 4x or more; (2) improved diversity through large-scale off-policy sampling; and (3) scalable search for sparse reward settings. On mathematical reasoning, preference-tuning, and automated red-teaming (diverse and representative post-training tasks), TBA produces speed and performance improvements over strong baselines.