Qualcomm（高通）

We present for video understanding (classification) a split space-time attention model, VideoSEMA, consisting of a scalable and efficient Mamba-like attention (SEMA) block in space and a softmax temporal attention in time. In each frame, SEMA attention applies a local window attention in parallel with a global averaging in a Mamba macro-architecture, which is called Mamba-like. Under certain rank conditions, we prove that the computationally cheaper split space-time attention is equivalent to full space-time attention. On benchmark K400 data sets, VideoSEMA out-performs heavier vision transformer and Mamba models. On benchmark SSv2 data, VideoSEMA leads in top-1 accuracy among models of similar parameter sizes. As image resolution scales up from standard $224^2$ to $1024^2$ on K400 and without fine-tuning, VideoSEMA degrades much more gracefully than VideoMamba in accuracy. It is promising to extend VideoSEMA to longer videos with a dilated/sparse temporal attention.

Attention is the critical component of a transformer. Yet the quadratic computational complexity of vanilla full attention in the input size and the inability of its linear attention variant to focus have been challenges for computer vision tasks. We provide a mathematical definition of generalized attention and formulate both vanilla softmax attention and linear attention within the general framework. We prove that generalized attention disperses, that is, as the number of keys tends to infinity, the query assigns equal weights to all keys. Motivated by the dispersion property and recent development of Mamba form of attention, we design Scalable and Efficient Mamba like Attention (SEMA) which utilizes token localization to avoid dispersion and maintain focusing, complemented by theoretically consistent arithmetic averaging to capture global aspect of attention. We support our approach on Imagenet-1k where classification results show that SEMA is a scalable and effective alternative beyond linear attention, outperforming recent vision Mamba models on increasingly larger scales of images at similar model parameter sizes.

The $\mathcal{O}(N^2)$ complexity of attention over $N$ tokens remains a computational bottleneck in transformer models. Vector-Quantized (VQ) attention reduces this to $\mathcal{O}(MN)$ by representing keys with $M$ codewords, but applies uniform codebook capacity regardless of where attention mass concentrates: high-attention regions of key space may be coarsely approximated while low-attention regions waste representational capacity. We propose Adaptive Vector-Quantized (AVQ) Attention, which adaptively allocates codebook capacity based on attention importance. Starting from a small set of codewords, our method identifies the most important codes during the forward pass and refines them with pre-learned child codewords, achieving fine-grained quantization where it matters most while maintaining coarse quantization elsewhere. We develop an implementation using custom Triton kernels that enables the full adaptive refinement process, including importance scoring, child codeword insertion, and parent contribution replacement, to be carried out within the tiled computation paradigm of Flash Attention with minimal overhead. Our approach maintains $\mathcal{O}(MN)$ complexity while achieving improved accuracy-efficiency trade-offs compared to fixed-codebook VQ-attention.

Autonomous driving operates in partially observable environments where actors may become fully occluded by other vehicles or infrastructure. Most end-to-end driving systems implicitly couple actor existence to instantaneous observations, causing actor hypotheses to degrade or disappear during prolonged occlusion and removing potentially critical agents from downstream prediction and planning. We introduce BeyondSight, a permanence-aware end-to-end driving framework that decouples actor existence from observability by maintaining persistent actor hypotheses over time. BeyondSight propagates actor queries temporally and updates them with observation-conditioned evidence, enabling joint perception, prediction, and planning to reason about actors even when they are temporarily unobservable. To enable principled training and evaluation of persistence-aware models, we further introduce nuScenes-Permanence, an extension of nuScenes that provides supervision and observability-conditioned evaluation for unobservable actors. Experiments show that BeyondSight substantially improves reasoning under occlusion, increasing detection performance for unobservable actors from 0 to 0.249 mAP while reducing planning error from 0.61 to 0.54 L2avg. These results highlight object permanence as an important modeling principle for robust end-to-end autonomous driving.

A striking feature of the human visual system is that it ingests visual information through a series of local foveated glimpses, rather than a single global computation. This makes human vision distinctly different from most popular computer vision models in use today, which input images globally and in a single shot. A natural question therefore is whether local, sequential vision models may provide any fundamental computational benefits in addition to being biologically more plausible than global models. In this work, we investigate this question from the perspective of visual state tracking and length generalization. Inspired by recent studies of length generalization in language models, we study the behavior of vision models trained on simple vision tasks that require the aggregation of local information across an image. Our experiments reveal that, similar to language models, vision models can learn to exploit global shortcuts and thereby fail to generalize over task length or complexity. We also show that recurrent vision policies based on strictly local perception can mitigate these failures, thereby allowing models to generalize on these tasks. Our results show that local attention may be an essential overlooked requirement for robust compositional generalization.

The quadratic cost of causal self-attention severely bottlenecks long-context transformer inference. While numerous post hoc linearization pipelines exist, it is difficult to identify which components preserve model quality. This work isolates the effect of state update design in a strict frozen-backbone regime. We show that softmax relies on key-dependent, rank-1 orthogonal projections, elucidating why delta-style networks outperform purely gated accumulation. We identify a potential source of approximation errors and introduce structural interventions, specifically sink tokens, short convolutions, and fixed-budget cache routing, which reduces the remaining gap. We scale this linearization approach across LLaMA and Qwen models up to 32B parameters, outperforming prior post hoc baselines on MMLU and matching the long-context retrieval of complex adaptive-caching frameworks.

Large language models (LLMs) require substantial compute, and thus energy, at inference time. While quantizing weights and activations is effective at improving efficiency, naive quantization of LLMs can significantly degrade performance due to large magnitude outliers. This paper describes FPTQuant, which introduces three novel, lightweight, and expressive function-preserving transforms (FPTs) to facilitate quantization of transformers: (1) a mergeable pre-RoPE transform for queries and keys, (2) a mergeable transform for values, and (3) a cheap, dynamic per-token scaling transform. By leveraging the equivariances and independencies inherent to canonical transformer operation, we designed these FPTs to maintain the model's function while shaping the intermediate activation distributions to be more quantization friendly. FPTQuant requires no custom kernels and adds virtually no overhead during inference. The FPTs are trained both locally to reduce outliers, and end-to-end such that the outputs of the quantized and full-precision models match. FPTQuant enables static INT4 quantization with minimal overhead and shows SOTA speed-up of up to 3.9X over FP. Empirically, FPTQuant has an excellent accuracy-speed trade-off -- it is performing on par or exceeding most prior work and only shows slightly lower accuracy compared to a method that is up to 29% slower.

End-to-end autonomous driving (E2E-AD) aims to directly map raw sensor information to driving actions. Recently, with the rapid advancement of multi-modal large language models (MLLMs), researchers have proposed the paradigm of Vision-Language-Action (VLA) models for E2E-AD, where it seeks to integrate visual perception, language understanding and action prediction within a single policy. However, existing VLA-based policies largely adopts imitation learning, where it only learns to drive by optimizing distance-based metrics w.r.t. logged expert trajectories. Such distribution shift between open-loop training and closed-loop inference leads to suboptimal performance in closed-loop planning. To close this gap, we present CLEAR, a system that enables closed-loop training using Reinforcement Learning (RL) at scale for E2E-AD. We propose to learn a novel residual waypoint policy around the waypoint prior from pretrained VLA policies, effectively harnessing the knowledge within. On another front, one of the key challenges to scale up RL for vision-based policies is the number of parallel simulation environments since RL is data hungry. To that end, we design a heterogeneous pipeline that places the simulator and the VLA learner on distinct compute groups, which allows us to dramatically increase the number of simulation environments running in parallel while avoiding resource contention and maintaining training stability. We show that with a simple reward, CLEAR significantly outperforms previous methods and sets new state-of-the-art performance on the challenging benchmarks of CARLA longest6 v2 and Bench2Drive.

Modern image-and-text-to-video diffusion models can synthesize highly realistic videos by iteratively denoising an initial Gaussian noise tensor conditioned on reference image and text inputs. However, existing approaches still lack precise and unified controllability over both object motion and camera motion within a single generation process. We present UniCaMo, a unified framework that enables simultaneous control of object trajectories and camera viewpoints by directly constructing the input noise of the diffusion model. Specifically, UniCaMo builds a shared 3D-grounded motion-consistent noise space across latent video frames. Sparse 3D point tracks are used to warp the Gaussian noise of the reference frame along desired object trajectories, while a virtual spherical noise representation provides globally consistent noise values for newly revealed scene regions under camera motion. By combining local track-guided noise warping with global sphere-based noise sampling, UniCaMo maintains geometric and temporal consistency under both object movement and viewpoint changes. Because UniCaMo modifies only the input noise, it requires no auxiliary adapters, control branches, or architectural changes to the underlying video diffusion model. With lightweight LoRA fine-tuning on large pretrained video diffusion models, including Wan 2.1 (14B), UniCaMo achieves state-of-the-art results in both video quality and motion controllability on standard controllable video generation benchmarks.

Recent advances in 3D content generation from text or images have achieved impressive results, yet view inconsistency from 2D generators and the scarcity of high-quality 3D data remain significant bottlenecks. Existing solutions typically adapt large-scale pre-trained text-to-image latent diffusion models to generate 3D Gaussian Splats (3DGS). However, these approaches often rely on training complex cascade pipelines that are computationally expensive and scalability-limited. Most critically, the quality of generated 3D assets is inherently constrained by each component capacity and compressed latent space, leading to decoding artifacts and accumulated errors. To address these limitations, we propose PixGS, a single-stage pipeline for direct high-quality 3DGS generation, which leverages recent advances in pixel-space diffusion to bypass lossy latent compression while still benefiting from the vast 2D generative priors. By directly denoising 3D Gaussian attributes at each timestep, our method enables precise, splat-level regularization of both appearance and geometry. Furthermore, we introduce a comprehensive supervision strategy that incorporates surface normals, depth, and high-frequency structural information, which is often overlooked in prior works. Experiments demonstrate that PixGS outperforms current state-of-the-art methods while maintaining a fast inference speed (1s on a single A100 GPU), offering a robust and efficient alternative to multi-stage generation pipelines.

In this paper, we introduce BamiBERT, a new BERT-based pre-trained language model for Vietnamese that addresses key limitations of PhoBERT -- the current de facto Vietnamese text encoder. Trained from scratch on a 129GB corpus of general-domain Vietnamese text for 20 epochs, BamiBERT supports an extended context length of up to 2048 tokens and operates directly on raw input, eliminating the need for external word segmentation. Across 8 Vietnamese benchmarks, it achieves the best score on 11 of 15 metrics and the second-best on 3 others, setting a new state of the art among "base"-sized Vietnamese encoders and demonstrating strong cross-domain generalization. We release BamiBERT at: https://huggingface.co/Qualcomm-AI-Research/BamiBERT

Despite decades of work, surveillance still struggles in searching and reasoning about specific targets across long, multi-camera videos. Existing methods - tracking, retrieval, and video LLMs require heavy manual filtering, capture only shallow attributes, and fail at temporal understanding. Prior benchmarks are also limited to basic retrieval and question answering, without addressing real world challenges that often involve multimodal queries and temporal grounding (e.g., "When did this person join the fight?" with the person's image). To address this gap, we introduce ForeSeaQA, a new benchmark specifically designed for video QA with image-and-text queries and timestamped annotations of key events. The dataset consists of long-horizon surveillance footage paired with diverse multimodal questions, enabling systematic evaluation of retrieval, temporal grounding, and multimodal reasoning in realistic forensic conditions. Not limited to this benchmark, we propose ForeSea, an AI forensic search system with a 3-stage, plug-and-play pipeline. (1) A tracking module filters irrelevant footage; (2) a multimodal embedding module indexes the remaining clips; and (3) during inference, the system retrieves top-K candidate clips for a video LLM to answer queries and localize events. On ForeSeaQA benchmark, ForeSea improves accuracy by 3.1 points and temporal IoU by 10.1 points over prior retrieval-augmented baselines. To our knowledge, ForeSeaQA is the first benchmark to support complex multimodal queries with precise temporal grounding, and ForeSea is the first VideoRAG system built to excel in this setting.

Modern one-step diffusion models achieve impressive quality through distribution-based timestep distillation. Yet, they rely on a critical assumption: Teacher and Student must inhabit the same latent space. This Shared-Space constraint prevents knowledge transfer from modern high-capacity Teachers (e.g., SD 3.5 and Flux) into compact, deployment-friendly Students such as SD 1.5, whose latent resolution and VAE parameterization differ from the Teacher. We formalize this overlooked regime as Cross-Space Distillation, where Teacher and Student differ in both latent resolution and VAE space. To enable distillation under this mismatch, we introduce the Bridge, a lightweight latent interface that maps Student latents into the Teacher space without modifying the Student backbone. Bridge combines a frozen Student VAE decoder as a spatial prior with a compact learnable projector, and is trained with latent reconstruction and attention fidelity objectives for stable Teacher-space alignment. Across diverse modern Teachers, Bridge enables substantial gains for compact one-step Students; for example, it improves SD 1.5 from 5.4 to 9.4 HPSv3 while preserving one-step inference, low latency, and broad ecosystem compatibility. These results show that heterogeneous large Teachers can be distilled into efficient, deployable backbones through a lightweight latent-space interface.

State-based fine-tuning has emerged as a compelling alternative to weight-based adaptation for transformers, updating lightweight controls into states rather than model weights, offering substantial memory savings while retaining parameter efficiency. However, most existing state-based methods typically apply only per-block control updates, which limits inter-block information exchange and restricts representational adaptation. Meanwhile, prior mechanisms that enable cross-block communication often introduce considerable computational overhead, reducing their practicality for efficient fine-tuning. We introduce Mixture-of-Control (MoC), a lightweight fine-tuning framework that adaptively integrates local and global control signals to enhance representation learning. MoC treats block-wise control states as experts in a sparse mixture-of-experts process, enabling efficient communication across transformer blocks. Empirical results across diverse transformer-based benchmarks demonstrate that MoC outperforms state-based methods while maintaining a comparable memory and computational efficiency.

Motions of objects and scenes carry essential intelligence in video understanding, offering rich cues for interpreting dynamic settings and interactions. Due to the cost and scarcity of high-quality annotation or ground truth of pixel-wise optical flow, however, motion estimation models are typically trained in synthetic domains while deployed in real-world domains. Addressing synthetic-to-real domain generalization challenges has been crucial for developing practical solutions in diverse open-world use cases. This paper introduces SciFlow, a simple yet effective, network-agnostic, training-based approach that leverages self-supervised learning to generalize motion estimation across synthetic and open-world domains. Specifically, SciFlow imposes semantic interference from open-world images onto synthetic images during training, blending indomain features with cross-domain interference, which enables the network to adapt to the real-world domains. Additionally, SciFlow utilizes geometric consistency to ensure validity of the self-supervision. Our experiment results show that SciFlow not only significantly enhances model robustness amidst domain variations, but also remarkably enables synthetic-to-real domain generalization without requiring any ground truth in the open world.

Knowledge Distillation (KD) enables training smaller student models under the guidance of larger teacher models, and the widely adopted TRL library implements it. Yet, TRL treats both models symmetrically, missing opportunities to exploit their pronounced asymmetry in memory footprint, and communication requirements. This paper presents an HPC-aware methodology for KD that decouples teacher and student partitioning efficiently. Our approach achieves up to 67% higher samples-per-second than TRL by avoiding unnecessary teacher-model data structures and selecting the best split strategy. We combine vertical and horizontal partitioning of models, deriving an analytical expression that identifies the existence of inflection points between splitting regimes. These results showed that exploiting teacher--student asymmetry through topology-aware parallelism notably accelerated GKD training on production HPC clusters at our company

Vision-Language Models (VLMs) achieve strong multimodal performance but are costly to deploy, and post-training quantization often causes significant accuracy loss. Despite its potential, quantization-aware training for VLMs remains underexplored. We propose GRACE, a framework unifying knowledge distillation and QAT under the Information Bottleneck principle: quantization constrains information capacity while distillation guides what to preserve within this budget. Treating the teacher as a proxy for task-relevant information, we introduce confidence-gated decoupled distillation to filter unreliable supervision, relational centered kernel alignment to transfer visual token structures, and an adaptive controller via Lagrangian relaxation to balance fidelity against capacity constraints. Across extensive benchmarks on LLaVA and Qwen families, our INT4 models consistently outperform FP16 baselines (e.g., LLaVA-1.5-7B: 70.1 vs. 66.8 on SQA; Qwen2-VL-2B: 76.9 vs. 72.6 on MMBench), nearly matching teacher performance. Using real INT4 kernel, we achieve 3$\times$ throughput with 54% memory reduction. This principled framework significantly outperforms existing quantization methods, making GRACE a compelling solution for resource-constrained deployment. Code and data are available at: https://github.com/ForeverBlue816/GRACE.

Answering natural-language questions over multi-hour audio requires both event recognition and temporal grounding. Current large audio-language models perform well on short clips, but are limited by context length, query-time cost, and weak temporal localization. We present LA-RAG (Long Audio-Retrieval Augmented Generation), a structured framework that converts continuous audio into timestamped event records using an open-vocabulary Audio Grounding Model (AGM), stores them in a SQL event database, and answers queries through intent-aware retrieval followed by LLM-based generation. LA-RAG supports offline grounding mode, where long recordings are pre-indexed for low-latency QA, and inference-time grounding mode, where query-conditioned grounding is performed for shorter open-ended clips. We create 24-hour Home-IoT and Industrial-IoT audio benchmarks and augment CASTELLA, a real-world audio moment retrieval dataset with QA pairs. In offline grounding mode, LA-RAG achieves 76.88% overall accuracy on Home-IoT and 71.10% on Industrial-IoT, with average query latencies below 0.6 seconds. In inference-time grounding mode, state-of-the-art LALMs achieve competitive event-detection accuracy on CASTELLA-QA but low temporal detection F1. We further show that LALMs augmented with our structured retrieval metadata achieve consistent temporal detection improvements, with F1 gains of 11-17% across baseline models with improved latency. These results show that explicit timestamped grounding and structured retrieval provide a practical complement to generative audio-language models for deployment-oriented long-audio QA.

Leveraging Vision Foundation Models (VFMs) for camera-to-LiDAR knowledge distillation offers a promising solution to the scarcity of annotated data needed to represent the immense geometric and kinematic diversity of real-world autonomous driving (AD). However, current approaches typically treat VFMs as black-box teachers, relying exclusively on frame-wise feature similarity. Consequently, they do not fully exploit the teacher's layer-wise semantic structure and global context, as well as the rich spatiotemporal information inherent in LiDAR sequences. We propose HilDA, a self-supervised pretraining framework for LiDAR backbones that better captures the semantic what and geometric where needed for driving tasks. HilDA combines hierarchical distillation comprising multi-layer distillation for progressive semantic alignment and global context distillation for scene-level semantics, with a temporal occupancy diffusion objective promoting spatiotemporal consistency. Models pre-trained with HilDA achieve state-of-the-art results on cross-modal distillation benchmarks and outperform models trained via prior distillation approaches on 3D object detection, scene flow, and semantic occupancy prediction. Code available at: https://maxiuw.github.io/hilda.

Current audio captioning relies on supervised learning with paired audio-caption data, which is costly to curate and may not reflect human preferences in real-world scenarios. To address this, we propose a preference-aligned audio captioning framework based on Reinforcement Learning from Human Feedback (RLHF). To capture nuanced preferences, we train a Contrastive Language-Audio Pretraining (CLAP) based reward model using human-labeled pairwise preference data. This reward model is integrated into an RL framework to fine-tune any baseline captioning system without ground-truth annotations. Extensive human evaluations across multiple datasets show that our method produces captions preferred over baseline models, particularly when baselines fail to provide correct and natural captions. Furthermore, our framework achieves performance comparable to supervised approaches with ground-truth data, demonstrating effective alignment with human preferences and scalability in real-world use.

Early lane-change intention prediction is essential for autonomous driving and ADAS, but it remains challenging because lane-changing behavior depends on evolving traffic risk, surrounding-vehicle interactions, and target-lane feasibility rather than only instantaneous vehicle states. This study proposes an evolutionary physics-informed temporal fusion framework for three-class lane-change intention prediction, including left lane change, right lane change, and no lane change. Instead of using static physics-informed variables alone, the proposed method derives temporal descriptors from conventional traffic signals, including risk evolution, gap persistence, counterfactual lane utility, interaction pressure gradient, maneuver feasibility, and intent consistency. These descriptors are fused with temporal embeddings learned from raw trajectory sequences through a sequence encoder, and the fused representation is used for final classification. Experiments are conducted on the highD and exiD datasets under 1\,s, 2\,s, and 3\,s prediction horizons. The proposed model achieves Macro F1-scores of 0.9514, 0.9256, and 0.8872 on highD, and 0.9386, 0.9070, and 0.8531 on exiD, respectively. The improvement is especially pronounced in exiD ramp-adjacent scenarios, indicating that temporal physical evolution is particularly useful in interaction-rich environments. These results demonstrate that combining evolutionary physics-informed descriptors with learned temporal representations provides a more dynamic and interpretable solution for early lane-change intention prediction.

Learning everyday skills, like cooking a dish, relies increasingly on instructional media such as online videos. This opens the door to the use of video (and multimodal) large language models (LLMs) as task guidance assistants. A crucial capability for the real-world success of a prospective task guidance assistant is it's ability to intervene proactively as soon as a mistake is apparent in order to guide the user. To evaluate this crucial capability, we introduce Ego-MC-Bench (Mistake Corrections), a benchmark for evaluating reactive, step-by-step task guidance in realistic cooking scenarios. Extensive experiments show that Ego-MC-Bench is highly challenging for state-of-the-art video LLMs. We argue that a key reason is the limited availability of training data for fine-tuning models on this task. Although there exists a wide range of cooking video datasets, existing datasets lack examples of mistakes along with appropriately timed interventions. To help address this data limitation, we also introduce Ego-CoMist, a counterfactual synthetic dataset created by transforming non -interactive cooking videos into supervised training examples showing proactive interventions. We show that fine-tuning on Ego-CoMist yields performance gains especially for smaller and more efficient video LLMs that are well suited for delivering assistance on edge devices.

Motion forecasting enables autonomous vehicles to anticipate scene evolution by predicting the future trajectories of dynamic agents. However, existing approaches typically assume a closed-world setting with a fixed object taxonomy and access to high-quality perception, limiting their applicability in the real world where perception is imperfect, and new object classes may emerge over time. In this work, we introduce class-incremental motion forecasting, a novel setting in which new object classes are sequentially introduced over time and future object trajectories are predicted directly from camera images. We propose the first end-to-end framework for this setting, which adapts to newly introduced classes while mitigating catastrophic forgetting of previously learned ones. Our method generates motion forecasting pseudo-labels for known classes and matches them with 2D instance masks from an open-vocabulary segmentation model. This 3D-to-2D keypoint voting mechanism filters inconsistent and overconfident predictions, while a query feature variance-based replay strategy samples informative past sequences to preserve prior knowledge. Extensive evaluations on nuScenes and Argoverse 2 show that our approach successfully preserves performance on known classes while effectively adapting to novel ones. We further demonstrate zero-shot transfer to real-world driving and show that the framework extends naturally to open- and closed-loop end-to-end class-incremental planning on nuScenes and NeuroNCAP. Code and models will be made publicly available at https://omen.cs.uni-freiburg.de.

Real-time conversational assistants for procedural manual tasks often depend on video input, which can be computationally expensive and compromise user privacy. For the first time, we propose a real-time conversational assistant that provides comprehensive guidance for procedural manual tasks using only lightweight privacy-preserving modalities such as audio and IMU inputs from a user's wearable device to understand the context. Using a furniture assembly task and a cooking task, we show how this assistant proactively communicates step-by-step instructions to a user performing a procedural task, and answers user questions. We illustrate the data generation method and the system design to achieve such an assistant. On observing that an off-the-shelf language model is a talkative assistant but is not always able to answer questions correctly, we demonstrate how finetuning the model improves its ability to limit unnecessary dialogues with a 50% increase in the precision, while also improving its ability to answer questions correctly, measured by a 150% increase in the recall of answers. We further describe how such an assistant is implemented on an edge device with no dependence on the cloud.

Test-time scaling via sequential revision has emerged as a powerful paradigm for enhancing Large Language Model (LLM) reasoning. However, standard post-training methods primarily optimize single-shot objectives, creating a fundamental misalignment with multi-step inference dynamics. While recent work treats this as multi-turn reinforcement learning (RL), conventional approaches optimize over the multi-step trajectories directly, failing to further exploit the high-quality mistakes in intermediate steps that model can learn from correcting them. We propose a two-stage iterative framework that alternates between online data/prompt augmentation and policy optimization. By converting the intermediate steps (``near-miss'' answers) in the successful recovery trajectories into decoupled revision and verification prompts, our approach concentrates training on both effective answer transformation and error identification. This approach enables efficient off-policy data generation and reduces the computational overhead of long-horizon sampling compared to standard multi-turn RL. On LiveCodeBench, using publicly available test cases as feedback, we observe gains of +6.5 points over the RL baseline and +4.0 points over standard multi-turn training. Beyond coding, our approach matches the previously reported SOTA result on circle packing while using the smallest base model (4B) and far fewer rollouts than the much larger evolutionary search systems. Math results under ground-truth verification further confirm improved correction ability. It also generalizes to out-of-distribution constraint-satisfaction puzzles such as n\_queens and mini\_sudoku, where correctness is defined entirely by problem constraints. Code is available at https://github.com/yxliu02/REVES.git.

Multimodal large language models have made rapid progress in video temporal grounding, yet real-world applications routinely require localizing every event that satisfies compositional temporal and spatial conditions. Existing benchmarks fall short: they localize only a single moment per query, count without temporal conditions, or treat grounding and counting as disjoint tasks. We introduce CoMET-Bench for Conditional Multi-Event Temporal Grounding in long-form video, comprising 2789 queries over 600 videos averaging 33.8 minutes across five real-world domains, with each query composed from 4 temporal conditions, 3 spatial conditions, and a dedicated negative-query subset. We further propose a unified evaluation protocol jointly measuring counting, grounding, and negative-query recognition, including a new Rejection-F1 metric that prevents trivial gaming by lazy "always-empty" models. Benchmarking a broad suite of MLLMs, agent-based, and grounding-specialized methods reveals that existing approaches remain far from solving this task. Building on these findings, we propose CoMET-Agent, a training-free agentic framework that reformulates the task as structured search-and-aggregate, improving F1@0.5 by 6.1% over GPT-5 purely through structural reasoning. Failure analysis further surfaces three open directions: fine-grained entity tracking, position-uniform retrieval, and causal event pairing.

While moral reasoning has emerged as a promising research direction for large language models (LLMs), achieving robust generalization remains a critical challenge. This challenge arises from the gap between what is said and what is morally implied. In this paper, we build on metapragmatic links and Moral Foundations Theory to close this gap. Specifically, we develop a pragmatic inference approach that enables LLMs, given a moral situation, to acquire the metapragmatic links between moral reasoning objectives and the social variables that influence them. We adapt this approach to three different moral reasoning tasks to demonstrate its adaptability and generalizability. Experimental results show that our approach significantly enhances LLMs' generalization in moral reasoning, paving the way for future research to leverage pragmatic inference across a wide range of moral reasoning tasks.

Retrieval-Augmented Generation (RAG) pipelines are compute-intensive, combining embedding, retrieval, reranking, and large language model (LLM) generation. Running them entirely on-device benefits privacy, latency, and offline use, but the energy cost of CPU inference is a major barrier. We present what is, to our knowledge, the first end-to-end RAG pipeline that runs all neural stages -- embedding, reranking, and LLM generation -- on the Qualcomm Hexagon NPU of the Snapdragon X Elite. Profiling on a Dell XPS 13 laptop, we compare NPU-accelerated RAG against CPU and OpenCL/Adreno GPU baselines on indexing and query workloads. On indexing, the NPU achieves 9.1x higher embedding throughput and 12.3x less system energy. On a 120-query Wikipedia-passage benchmark, it delivers 18.1x faster LLM prefilling, 4.0x lower end-to-end query latency, and 4.0x less system energy than the CPU baseline; the same workload on the integrated GPU is 1.7x slower than CPU and uses 6.5x more energy than the NPU. A GPT-4.1 LLM-as-judge evaluation finds NPU answer quality on par with CPU and GPU within evaluator noise (mean 9.32 vs. 8.95 vs. 9.03 on a 1-10 rubric), with 86.7% of queries scoring identically across all three backends. On the Snapdragon X Elite / Hexagon class of laptop SoC, the NPU thus enables practical, energy-efficient on-device RAG without quality regression -- a sustainable path toward green edge intelligence that we expect to generalize to comparable mobile NPUs (Apple Neural Engine, Intel NPU, MediaTek APU) as their software stacks mature.

Deploying safe large language models (LLMs) on resource-constrained edge devices presents a critical challenge: while dual-model systems combining LLMs with guard models provide effective safety guarantees, their substantial memory and computational demands make them prohibitively expensive for on-device deployment. This paper presents a comprehensive study of parameter-efficient safety alignment methods for resource-constrained settings. Through systematic evaluation across multiple LLM architectures, training objectives, and parameter-efficient fine-tuning approaches, we identify that soft prompts combined with distillation-based training consistently outperform alternative methods. We introduce distillation frameworks based on total variation and KL divergence that effectively transfer safety behaviors from guard models into learned soft prompts. Our evaluations on various benchmarks demonstrate that this combination achieves superior safety-usefulness trade-offs compared to LoRA adapters, steering vectors, and direct optimization methods, while requiring minimal additional memory and compute at inference time. These findings establish soft prompt distillation as the preferred approach for safety alignment in on-device LLM deployment.

Autoregressive (AR) image generation models are highly expressive but computationally intensive, motivating effective model compression. Knowledge distillation (KD) is a natural approach for model compression and has been widely studied in language modeling, yet its behavior in visual AR generation remains underexplored. In this work, we present the first systematic study of distillation strategies for AR image models. Our analysis shows that while standard distillation can yield meaningful gains, recent methods developed for language do not directly transfer to images: long decoding horizons and visual token ambiguity make teacher supervision unreliable especially under student-conditioned contexts. To address this, we propose VarKD, a distillation framework for visual autoregressive models that distills on student samples while selectively applying teacher supervision and reducing token-level ambiguity. Experiments on ImageNet across multiple AR backbones show that VarKD consistently outperforms prior distillation baselines, narrowing the gap to large-scale models.