arXiv subjects

Dinesh Manocha

Publications and source records attributed to Dinesh Manocha.

At least 91 records · Page 5Linked to original sources

arXiv2026

Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away

Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But recent evidence shows that it can simultaneously degrade safety alignment and increase jailbreak success rates. We propose SafeThink, a lightweight inference-time defense that treats safety recovery as a satisficing constraint rather than a maximization objective. SafeThink monitors the evolving reasoning trace with a safety reward model and conditionally injects an optimized short corrective prefix ("Wait, think safely") only when the safety threshold is violated. In our evaluations across six open-source MLRMs and four jailbreak benchmarks (JailbreakV-28K, Hades, FigStep, and MM-SafetyBench), SafeThink reduces attack success rates by 30-60% (e.g., LlamaV-o1: 63.33% to 5.74% on JailbreakV-28K, R1-Onevision: 69.07% to 5.65% on Hades) while preserving reasoning performance (MathVista accuracy: 65.20% to 65.00%). A key empirical finding from our experiments is that safety recovery is often only a few steering steps away: intervening in the first 1-3 reasoning steps typically suffices to redirect the full generation toward safe completions.

cs.CL↗

arXiv2026

EgoAVU: Egocentric Audio-Visual Understanding

Understanding egocentric videos plays a vital role for embodied intelligence. Recent multi-modal large language models (MLLMs) can accept both visual and audio inputs. However, due to the challenge of obtaining text labels with coherent joint-modality information, whether MLLMs can jointly understand both modalities in egocentric videos remains under-explored. To address this problem, we introduce EgoAVU, a scalable data engine to automatically generate egocentric audio-visual narrations, questions, and answers. EgoAVU enriches human narrations with multimodal context and generates audio-visual narrations through cross-modal correlation modeling. Token-based video filtering and modular, graph-based curation ensure both data diversity and quality. Leveraging EgoAVU, we construct EgoAVU-Instruct, a large-scale training dataset of 3M samples, and EgoAVU-Bench, a manually verified evaluation split covering diverse tasks. EgoAVU-Bench clearly reveals the limitations of existing MLLMs: they bias heavily toward visual signals, often neglecting audio cues or failing to correspond audio with the visual source. Finetuning MLLMs on EgoAVU-Instruct effectively addresses this issue, enabling up to 113% performance improvement on EgoAVU-Bench. Such benefits also transfer to other benchmarks such as EgoTempo and EgoIllusion, achieving up to 28% relative performance gain. Code will be released to the community.

cs.CV↗

arXiv2026

MemCtrl: Using MLLMs as Active Memory Controllers on Embodied Agents

Foundation models rely on in-context learning for personalized decision making. The limited size of this context window necessitates memory compression and retrieval systems like RAG. These systems however often treat memory as large offline storage spaces, which is unfavorable for embodied agents that are expected to operate under strict memory and compute constraints, online. In this work, we propose MemCtrl, a novel framework that uses Multimodal Large Language Models (MLLMs) for pruning memory online. MemCtrl augments MLLMs with a trainable memory head μthat acts as a gate to determine which observations or reflections to retain, update, or discard during exploration. We evaluate with training two types of μ, 1) via an offline expert, and 2) via online RL, and observe significant improvement in overall embodied task completion ability on μ-augmented MLLMs. In particular, on augmenting two low performing MLLMs with MemCtrl on multiple subsets of the EmbodiedBench benchmark, we observe that μ-augmented MLLMs show an improvement of around 16% on average, with over 20% on specific instruction subsets. Finally, we present a qualitative analysis on the memory fragments collected by μ, noting the superior performance of μaugmented MLLMs on long and complex instruction types.

cs.AI↗

arXiv2025

AMUSE: Audio-Visual Benchmark and Alignment Framework for Agentic Multi-Speaker Understanding

Recent multimodal large language models (MLLMs) such as GPT-4o and Qwen3-Omni show strong perception but struggle in multi-speaker, dialogue-centric settings that demand agentic reasoning tracking who speaks, maintaining roles, and grounding events across time. These scenarios are central to multimodal audio-video understanding, where models must jointly reason over audio and visual streams in applications such as conversational video assistants and meeting analytics. We introduce AMUSE, a benchmark designed around tasks that are inherently agentic, requiring models to decompose complex audio-visual interactions into planning, grounding, and reflection steps. It evaluates MLLMs across three modes zero-shot, guided, and agentic and six task families, including spatio-temporal speaker grounding and multimodal dialogue summarization. Across all modes, current models exhibit weak multi-speaker reasoning and inconsistent behavior under both non-agentic and agentic evaluation. Motivated by the inherently agentic nature of these tasks and recent advances in LLM agents, we propose RAFT, a data-efficient agentic alignment framework that integrates reward optimization with intrinsic multimodal self-evaluation as reward and selective parameter adaptation for data and parameter efficient updates. Using RAFT, we achieve up to 39.52\% relative improvement in accuracy on our benchmark. Together, AMUSE and RAFT provide a practical platform for examining agentic reasoning in multimodal models and improving their capabilities.

cs.AI↗

arXiv2025

SyncTrack4D: Cross-Video Motion Alignment and Video Synchronization for Multi-Video 4D Gaussian Splatting

Modeling dynamic 3D scenes is challenging due to their high-dimensional nature, which requires aggregating information from multiple views to reconstruct time-evolving 3D geometry and motion. We present a novel multi-video 4D Gaussian Splatting (4DGS) approach designed to handle real-world, unsynchronized video sets. Our approach, SyncTrack4D, directly leverages dense 4D track representation of dynamic scene parts as cues for simultaneous cross-video synchronization and 4DGS reconstruction. We first compute dense per-video 4D feature tracks and cross-video track correspondences by Fused Gromov-Wasserstein optimal transport approach. Next, we perform global frame-level temporal alignment to maximize overlapping motion of matched 4D tracks. Finally, we achieve sub-frame synchronization through our multi-video 4D Gaussian splatting built upon a motion-spline scaffold representation. The final output is a synchronized 4DGS representation with dense, explicit 3D trajectories, and temporal offsets for each video. We evaluate our approach on the Panoptic Studio and SyncNeRF Blender, demonstrating sub-frame synchronization accuracy with an average temporal error below 0.26 frames, and high-fidelity 4D reconstruction reaching 26.3 PSNR scores on the Panoptic Studio dataset. To the best of our knowledge, our work is the first general 4D Gaussian Splatting approach for unsynchronized video sets, without assuming the existence of predefined scene objects or prior models.

cs.CV↗

arXiv2025

MoRe: Monocular Geometry Refinement via Graph Optimization for Cross-View Consistency

Monocular 3D foundation models offer an extensible solution for perception tasks, making them attractive for broader 3D vision applications. In this paper, we propose MoRe, a training-free Monocular Geometry Refinement method designed to improve cross-view consistency and achieve scale alignment. To induce inter-frame relationships, our method employs feature matching between frames to establish correspondences. Rather than applying simple least squares optimization on these matched points, we formulate a graph-based optimization framework that performs local planar approximation using the estimated 3D points and surface normals estimated by monocular foundation models. This formulation addresses the scale ambiguity inherent in monocular geometric priors while preserving the underlying 3D structure. We further demonstrate that MoRe not only enhances 3D reconstruction but also improves novel view synthesis, particularly in sparse view rendering scenarios.

cs.CV↗

arXiv2025

PhysGS: Bayesian-Inferred Gaussian Splatting for Physical Property Estimation

Understanding physical properties such as friction, stiffness, hardness, and material composition is essential for enabling robots to interact safely and effectively with their surroundings. However, existing 3D reconstruction methods focus on geometry and appearance and cannot infer these underlying physical properties. We present PhysGS, a Bayesian-inferred extension of 3D Gaussian Splatting that estimates dense, per-point physical properties from visual cues and vision--language priors. We formulate property estimation as Bayesian inference over Gaussian splats, where material and property beliefs are iteratively refined as new observations arrive. PhysGS also models aleatoric and epistemic uncertainties, enabling uncertainty-aware object and scene interpretation. Across object-scale (ABO-500), indoor, and outdoor real-world datasets, PhysGS improves accuracy of the mass estimation by up to 22.8%, reduces Shore hardness error by up to 61.2%, and lowers kinetic friction error by up to 18.1% compared to deterministic baselines. Our results demonstrate that PhysGS unifies 3D reconstruction, uncertainty modeling, and physical reasoning in a single, spatially continuous framework for dense physical property estimation. Additional results are available at https://samchopra2003.github.io/physgs.

cs.CV↗

arXiv2025

HACL: History-Aware Curriculum Learning for Fast Locomotion

We address the problem of agile and rapid locomotion, a key characteristic of quadrupedal and bipedal robots. We present a new algorithm that maintains stability and generates high-speed trajectories by considering the temporal aspect of locomotion. Our formulation takes into account past information based on a novel history-aware curriculum Learning (HACL) algorithm. We model the history of joint velocity commands with respect to the observed linear and angular rewards using a recurrent neural net (RNN). The hidden state helps the curriculum learn the relationship between the forward linear velocity and angular velocity commands and the rewards over a given time-step. We validate our approach on the MIT Mini Cheetah,Unitree Go1, and Go2 robots in a simulated environment and on a Unitree Go1 robot in real-world scenarios. In practice, HACL achieves peak forward velocity of 6.7 m/s for a given command velocity of 7m/s and outperforms prior locomotion algorithms by nearly 20%.

cs.RO↗

arXiv2025

3D-free meets 3D priors: Novel View Synthesis from a Single Image with Pretrained Diffusion Guidance

Recent 3D novel view synthesis (NVS) methods often require extensive 3D data for training, and also typically lack generalization beyond the training distribution. Moreover, they tend to be object centric and struggle with complex and intricate scenes. Conversely, 3D-free methods can generate text-controlled views of complex, in-the-wild scenes using a pretrained stable diffusion model without the need for a large amount of 3D-based training data, but lack camera control. In this paper, we introduce a method capable of generating camera-controlled viewpoints from a single input image, by combining the benefits of 3D-free and 3D-based approaches. Our method excels in handling complex and diverse scenes without extensive training or additional 3D and multiview data. It leverages widely available pretrained NVS models for weak guidance, integrating this knowledge into a 3D-free view synthesis style approach, along with enriching the CLIP vision-language space with 3D camera angle information, to achieve the desired results. Experimental results demonstrate that our method outperforms existing models in both qualitative and quantitative evaluations, achieving high-fidelity, consistent novel view synthesis at desired camera angles across a wide variety of scenes while maintaining accurate, natural detail representation and image clarity across various viewpoints. We also support our method with a comprehensive analysis of 2D image generation models and the 3D space, providing a solid foundation and rationale for our solution.

cs.CV↗

arXiv2025

DR. Nav: Semantic-Geometric Representations for Proactive Dead-End Recovery and Navigation

We present DR. Nav (Dead-End Recovery-aware Navigation), a novel approach to autonomous navigation in scenarios where dead-end detection and recovery are critical, particularly in unstructured environments where robots must handle corners, vegetation occlusions, and blocked junctions. DR. Nav introduces a proactive strategy for navigation in unmapped environments without prior assumptions. Our method unifies dead-end prediction and recovery by generating a single, continuous, real-time semantic cost map. Specifically, DR. Nav leverages cross-modal RGB-LiDAR fusion with attention-based filtering to estimate per-cell dead-end likelihoods and recovery points, which are continuously updated through Bayesian inference to enhance robustness. Unlike prior mapping methods that only encode traversability, DR. Nav explicitly incorporates recovery-aware risk into the navigation cost map, enabling robots to anticipate unsafe regions and plan safer alternative trajectories. We evaluate DR. Nav across multiple dense indoor and outdoor scenarios and demonstrate an increase of 83.33% in accuracy in detection, a 52.4% reduction in time-to-goal (path efficiency), compared to state-of-the-art planners such as DWA, MPPI, and Nav2 DWB. Furthermore, the dead-end classifier functions

cs.RO↗

arXiv2025

SPUR: A Plug-and-Play Framework for Integrating Spatial Audio Understanding and Reasoning into Large Audio-Language Models

Spatial perception is central to auditory intelligence, enabling accurate understanding of real-world acoustic scenes and advancing human-level perception of the world around us. While recent large audio-language models (LALMs) show strong reasoning over complex audios, most operate on monaural inputs and lack the ability to capture spatial cues such as direction, elevation, and distance. We introduce SPUR, a lightweight, plug-in approach that equips LALMs with spatial perception through minimal architectural changes. SPUR consists of: (i) a First-Order Ambisonics (FOA) encoder that maps (W, X, Y, Z) channels to rotation-aware, listener-centric spatial features, integrated into target LALMs via a multimodal adapter; and (ii) SPUR-Set, a spatial QA dataset combining open-source FOA recordings with controlled simulations, emphasizing relative direction, elevation, distance, and overlap for supervised spatial reasoning. Fine-tuning our model on the SPUR-Set consistently improves spatial QA and multi-speaker attribution while preserving general audio understanding. SPUR provides a simple recipe that transforms monaural LALMs into spatially aware models. Extensive ablations validate the effectiveness of our approach.

eess.AS↗

arXiv2025

Music Flamingo: Scaling Music Understanding in Audio Language Models

We introduce Music Flamingo, a novel large audio-language model designed to advance music (including song) understanding in foundational audio models. While audio-language research has progressed rapidly, music remains challenging due to its dynamic, layered, and information-dense nature. Progress has been further limited by the difficulty of scaling open audio understanding models, primarily because of the scarcity of high-quality music data and annotations. As a result, prior models are restricted to producing short, high-level captions, answering only surface-level questions, and showing limited generalization across diverse musical cultures. To address these challenges, we curate MF-Skills, a large-scale dataset labeled through a multi-stage pipeline that yields rich captions and question-answer pairs covering harmony, structure, timbre, lyrics, and cultural context. We fine-tune an enhanced Audio Flamingo 3 backbone on MF-Skills and further strengthen multiple skills relevant to music understanding. To improve the model's reasoning abilities, we introduce a post-training recipe: we first cold-start with MF-Think, a novel chain-of-thought dataset grounded in music theory, followed by GRPO-based reinforcement learning with custom rewards. Music Flamingo achieves state-of-the-art results across 10+ benchmarks for music understanding and reasoning, establishing itself as a generalist and musically intelligent audio-language model. Beyond strong empirical results, Music Flamingo sets a new standard for advanced music understanding by demonstrating how models can move from surface-level recognition toward layered, human-like perception of songs. We believe this work provides both a benchmark and a foundation for the community to build the next generation of models that engage with music as meaningfully as humans do.

eess.AS↗

arXiv2025

Accelerated, Memory-Efficient Far-Field Scattering Computation with Monte Carlo SBR

We introduce a Monte Carlo integration-based Shooting and Bouncing Ray (SBR) algorithm for electromagnetic scattering, specifically targeting complex dielectric materials. Unlike traditional deterministic SBR methods, our approach is the first to reformulate the SBR integral equations using Monte Carlo techniques and advanced variance reduction strategies adapted from photorealistic rendering. This enables efficient, massively parallel computation on modern GPUs, resulting in up to a 10-15x reduction in memory usage and a 4x speed up in runtime, particularly for multilayer dielectric structures. Our method emphasizes high-energy propagation paths, efficiently capturing long multipath and interreflection effects. Verification on canonical 3D geometries and ISAR imaging of both conducting and dielectric representative aircraft models demonstrates that our Monte Carlo SBR achieves high accuracy while maintaining low noise, making it suitable for downstream imaging and analysis tasks.

cs.CE↗

arXiv2025

Sensible Agent: A Framework for Unobtrusive Interaction with Proactive AR Agents

Proactive AR agents promise context-aware assistance, but their interactions often rely on explicit voice prompts or responses, which can be disruptive or socially awkward. We introduce Sensible Agent, a framework designed for unobtrusive interaction with these proactive agents. Sensible Agent dynamically adapts both "what" assistance to offer and, crucially, "how" to deliver it, based on real-time multimodal context sensing. Informed by an expert workshop (n=12) and a data annotation study (n=40), the framework leverages egocentric cameras, multimodal sensing, and Large Multimodal Models (LMMs) to infer context and suggest appropriate actions delivered via minimally intrusive interaction modes. We demonstrate our prototype on an XR headset through a user study (n=10) in both AR and VR scenarios. Results indicate that Sensible Agent significantly reduces perceived interaction effort compared to voice-prompted baseline, while maintaining high usability and achieving higher preference.

cs.HC↗

arXiv2025

VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding

Vision-Language Models (VLMs) have achieved strong results in video understanding, yet a key question remains: do they truly comprehend visual content or only learn shallow correlations between vision and language? Real visual understanding, especially of physics and common sense, is essential for AI systems that interact with the physical world. Current evaluations mostly use real-world videos similar to training data, so high benchmark scores may not reflect real reasoning ability. To address this, we propose negative-control tests using videos that depict physically impossible or logically inconsistent events. We introduce VideoHallu, a synthetic dataset of physics- and commonsense-violating scenes generated with Veo2, Sora, and Kling. It includes expert-annotated question-answer pairs across four categories of violations. Tests of leading VLMs (Qwen-2.5-VL, Video-R1, VideoChat-R1) show that, despite strong results on benchmarks such as MVBench and MMVU, they often miss these violations, exposing gaps in visual reasoning. Reinforcement learning fine-tuning on VideoHallu improves recognition of such violations without reducing standard benchmark performance. Our data is available at https://github.com/zli12321/VideoHallu.git.

cs.CV↗

arXiv2025

Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek R1) have led to a popular belief that extending thinking traces using prompts like "Wait" or "Let me rethink" can improve performance. This raises a natural question: Does thinking more at test-time truly lead to better reasoning? To answer this question, we perform a detailed empirical study across models and benchmarks, which reveals a consistent pattern of initial performance improvements from additional thinking followed by a decline, due to "overthinking". To understand this non-monotonic trend, we consider a simple probabilistic model, which reveals that additional thinking increases output variance-creating an illusion of improved reasoning while ultimately undermining precision. Thus, observed gains from "more thinking" are not true indicators of improved reasoning, but artifacts stemming from the connection between model uncertainty and evaluation metric. This suggests that test-time scaling through extended thinking is not an effective way to utilize the inference thinking budget. Recognizing these limitations, we introduce an alternative test-time scaling approach, parallel thinking, inspired by Best-of-N sampling. Our method generates multiple independent reasoning paths within the same inference budget and selects the most consistent response via majority vote, achieving up to 20% higher accuracy compared to extended thinking. This provides a simple yet effective mechanism for test-time scaling of reasoning models.

cs.AI↗

arXiv2025

Engagement Undermines Safety: How Stereotypes and Toxicity Shape Humor in Language Models

Large language models are increasingly used for creative writing and engagement content, raising safety concerns about the outputs. Therefore, casting humor generation as a testbed, this work evaluates how funniness optimization in modern LLM pipelines couples with harmful content by jointly measuring humor, stereotypicality, and toxicity. This is further supplemented by analyzing incongruity signals through information-theoretic metrics. Across six models, we observe that harmful outputs receive higher humor scores which further increase under role-based prompting, indicating a bias amplification loop between generators and evaluators. Information-theoretic analyses show harmful cues widen predictive uncertainty and surprisingly, can even make harmful punchlines more expected for some models, suggesting structural embedding in learned humor distributions. External validation on an additional satire-generation task with human perceived funniness judgments shows that LLM satire increases stereotypicality and typically toxicity, including for closed models. Quantitatively, stereotypical/toxic jokes gain $10-21\%$ in mean humor score, stereotypical jokes appear $11\%$ to $28\%$ more often among the jokes marked funny by LLM-based metric and up to $10\%$ more often in generations perceived as funny by humans.

cs.CL↗

arXiv2025

EgoSocial: Benchmarking Proactive Intervention Ability of Omnimodal LLMs via Egocentric Social Interaction Perception

As AR/VR technologies become integral to daily life, there's a growing need for AI that understands human social dynamics from an egocentric perspective. However, current LLMs often lack the social awareness to discern when to intervene as AI assistant. This leads to constant, socially unaware responses that may disrupt natural conversation and negatively impact user focus. To address these limitations, we introduce EgoSocial, a large-scale egocentric dataset with 13,500 social video-question pairs, specifically designed to benchmark intervention in social interaction perception. We also present an in-depth analysis of current omnimodal LLMs (OLLMs) to assess their effectiveness in detecting diverse social contextual cues. Experiments show that OLLMs still struggle to detect the intervention timing (14.4% for Gemini 2.5 Pro). We also propose EgoSoD (EgoSocial Detection), an end-to-end method for robustly discerning social dynamics. Informed by our OLLM analysis, EgoSoD integrates multimodal contextual cues (e.g., audio and visual cues) into a social thinking graph, dynamically modeling participants and interactions. Our method proactively detects intervention timing and social interactions, precisely determining when to intervene. Our EgoSoD improves Phi-4 by 45.6% and Gemini 2.5 Pro by 9.9% on Intervention Timing performance, and improves Phi-4 by 20.4% and Gemini 2.5 Pro by 6.9% on overall Social Interaction performance. We will release the dataset and code soon.

cs.CV↗