Large Language Models (LLMs) have demonstrated significant strides across various information retrieval tasks, particularly as rerankers, owing to their strong generalization and knowledge-transfer capabilities acquired from extensive pretraining. In parallel, the rise of LLM-based chat interfaces has raised user expectations, encouraging users to pose more complex queries that necessitate retrieval by ``reasoning'' over documents rather than through simple keyword matching or semantic similarity. While some recent efforts have exploited reasoning abilities of LLMs for reranking such queries, considerable potential for improvement remains. In that regards, we introduce InsertRank, an LLM-based reranker that leverages lexical signals like BM25 scores during reranking to further improve retrieval performance. InsertRank demonstrates improved retrieval effectiveness on -- BRIGHT, a reasoning benchmark spanning 12 diverse domains, and R2MED, a specialized medical reasoning retrieval benchmark spanning 8 different tasks. We conduct an exhaustive evaluation and several ablation studies and demonstrate that InsertRank consistently improves retrieval effectiveness across multiple families of LLMs, including GPT, Gemini, and Deepseek models. %In addition, we also conduct ablation studies on normalization by varying the scale of the BM25 scores, and positional bias by shuffling the order of the documents. With Deepseek-R1, InsertRank achieves a score of 37.5 on the BRIGHT benchmark. and 51.1 on the R2MED benchmark, surpassing previous methods.

full paperreport(pdf)

Recent advances in large language models (LLMs), particularly those enhanced through reinforced post-training, have demonstrated impressive reasoning capabilities, as exemplified by models such as OpenAI o1 and DeepSeek-R1. However, these capabilities are predominantly benchmarked on domains like mathematical problem solving and code generation -- leaving open the question of whether such reasoning skills generalize to complex, real-world scenarios. In this paper, we introduce LocationReasoner, a benchmark designed to evaluate LLMs' reasoning abilities in the context of real-world site selection, where models must identify feasible locations by reasoning over diverse and complicated spatial, environmental, and logistical constraints. The benchmark comprises over 300 carefully crafted queries of varying difficulty levels, supported by a sandbox environment with in-house tools for constraint-based location search. Extensive evaluations reveal that state-of-the-art reasoning models offer limited improvement over their non-reasoning predecessors in real-world contexts, with even the latest OpenAI o4 model failing on 30% of site selection tasks. Moreover, agentic strategies such as ReAct and Reflexion often suffer from over-reasoning, leading to worse outcomes than direct code-generation prompting. With key limitations of LLMs in holistic and non-linear reasoning highlighted, we release LocationReasoner to foster the development of LLMs and agents capable of robust, grounded reasoning in real-world decision-making tasks. Codes and data for our benchmark are available at https://github.com/miho-koda/LocationReasoner.

full paperreport(pdf)

We introduce LogiPlan, a novel benchmark designed to evaluate the capabilities of large language models (LLMs) in logical planning and reasoning over complex relational structures. Logical relational reasoning is important for applications that may rely on LLMs to generate and query structured graphs of relations such as network infrastructure, knowledge bases, or business process schema. Our framework allows for dynamic variation of task complexity by controlling the number of objects, relations, and the minimum depth of relational chains, providing a fine-grained assessment of model performance across difficulty levels. LogiPlan encompasses three complementary tasks: (1) Plan Generation, where models must construct valid directed relational graphs meeting specified structural constraints; (2) Consistency Detection, testing models' ability to identify inconsistencies in relational structures; and (3) Comparison Question, evaluating models' capacity to determine the validity of queried relationships within a given graph. Additionally, we assess models' self-correction capabilities by prompting them to verify and refine their initial solutions. We evaluate state-of-the-art models including DeepSeek R1, Gemini 2.0 Pro, Gemini 2 Flash Thinking, GPT-4.5, GPT-4o, Llama 3.1 405B, O3-mini, O1, and Claude 3.7 Sonnet across these tasks, revealing significant performance gaps that correlate with model scale and architecture. Our analysis demonstrates that while recent reasoning-enhanced models show promising results on simpler instances, they struggle with more complex configurations requiring deeper logical planning.

full paperreport(pdf)

Recent advancements in multimodal large language models have successfully extended the Reason-Then-Respond paradigm to image-based reasoning, yet video-based reasoning remains an underdeveloped frontier, primarily due to the scarcity of high-quality reasoning-oriented data and effective training methodologies. To bridge this gap, we introduce DarkEventInfer and MixVidQA, two novel datasets specifically designed to stimulate the model's advanced video understanding and reasoning abilities. DarkEventinfer presents videos with masked event segments, requiring models to infer the obscured content based on contextual video cues. MixVidQA, on the other hand, presents interleaved video sequences composed of two distinct clips, challenging models to isolate and reason about one while disregarding the other. Leveraging these carefully curated training samples together with reinforcement learning guided by diverse reward functions, we develop VersaVid-R1, the first versatile video understanding and reasoning model under the Reason-Then-Respond paradigm capable of handling multiple-choice and open-ended question answering, as well as video captioning tasks. Extensive experiments demonstrate that VersaVid-R1 significantly outperforms existing models across a broad spectrum of benchmarks, covering video general understanding, cognitive reasoning, and captioning tasks.

full paperreport(pdf)

Recent advances in Large Language Models (LLMs) have driven their adoption in recommender systems through Retrieval-Augmented Generation (RAG) frameworks. However, existing RAG approaches predominantly rely on flat, similarity-based retrieval that fails to leverage the rich relational structure inherent in user-item interactions. We introduce LlamaRec-LKG-RAG, a novel single-pass, end-to-end trainable framework that integrates personalized knowledge graph context into LLM-based recommendation ranking. Our approach extends the LlamaRec architecture by incorporating a lightweight user preference module that dynamically identifies salient relation paths within a heterogeneous knowledge graph constructed from user behavior and item metadata. These personalized subgraphs are seamlessly integrated into prompts for a fine-tuned Llama-2 model, enabling efficient and interpretable recommendations through a unified inference step. Comprehensive experiments on ML-100K and Amazon Beauty datasets demonstrate consistent and significant improvements over LlamaRec across key ranking metrics (MRR, NDCG, Recall). LlamaRec-LKG-RAG demonstrates the critical value of structured reasoning in LLM-based recommendations and establishes a foundation for scalable, knowledge-aware personalization in next-generation recommender systems. Code is available at~\href{https://github.com/VahidAz/LlamaRec-LKG-RAG}{repository}.

full paperreport(pdf)

Natural Language Inference (NLI) is the task of determining whether a sentence pair represents entailment, contradiction, or a neutral relationship. While NLI models perform well on many inference tasks, their ability to handle fine-grained pragmatic inferences, particularly presupposition in conditionals, remains underexplored. In this study, we introduce CONFER, a novel dataset designed to evaluate how NLI models process inference in conditional sentences. We assess the performance of four NLI models, including two pre-trained models, to examine their generalization to conditional reasoning. Additionally, we evaluate Large Language Models (LLMs), including GPT-4o, LLaMA, Gemma, and DeepSeek-R1, in zero-shot and few-shot prompting settings to analyze their ability to infer presuppositions with and without prior context. Our findings indicate that NLI models struggle with presuppositional reasoning in conditionals, and fine-tuning on existing NLI datasets does not necessarily improve their performance.

full paperreport(pdf)

The Group Relative Policy Optimization (GRPO) algorithm has demonstrated considerable success in enhancing the reasoning capabilities of large language models (LLMs), as evidenced by DeepSeek-R1. However, the absence of intermediate supervision in GRPO frequently leads to inefficient exploration dynamics. A single error in a complex reasoning chain can invalidate the entire solution, resulting in abrupt reward vanishing and compromising training stability.To address these challenges, we propose MGRPO (Multi-layer GRPO). MGRPO operates in two layers: the first layer employs standard GRPO to generate an initial response. This response, along with the original query, is then fed into a second-layer GRPO process. This second layer is specifically trained to identify and correct errors in the initial response, effectively creating a self-correction loop. This mechanism provides implicit process-level supervision by rewarding successful error correction, without requiring an explicit, densely-annotated reward model. Experimental results on several mathematical reasoning benchmarks demonstrate that MGRPO significantly outperforms standard GRPO, achieving superior performance by fostering both reasoning and self-correction abilities.

full paperreport(pdf)

With the rise of long-context language models (LMs) capable of processing tens of thousands of tokens in a single pass, do multi-stage retrieval-augmented generation (RAG) pipelines still offer measurable benefits over simpler, single-stage approaches? To assess this question, we conduct a controlled evaluation for QA tasks under systematically scaled token budgets, comparing two recent multi-stage pipelines, ReadAgent and RAPTOR, against three baselines, including DOS RAG (Document's Original Structure RAG), a simple retrieve-then-read method that preserves original passage order. Despite its straightforward design, DOS RAG consistently matches or outperforms more intricate methods on multiple long-context QA benchmarks. We recommend establishing DOS RAG as a simple yet strong baseline for future RAG evaluations, pairing it with emerging embedding and language models to assess trade-offs between complexity and effectiveness as model capabilities evolve.

full paperreport(pdf)

Large language models (LLMs) have shown remarkable reasoning capabilities when trained with chain-of-thought (CoT) supervision. However, the long and verbose CoT traces, especially those distilled from large reasoning models (LRMs) such as DeepSeek-R1, significantly increase training costs during the distillation process, where a non-reasoning base model is taught to replicate the reasoning behavior of an LRM. In this work, we study the problem of CoT condensation for resource-efficient reasoning training, aimed at pruning intermediate reasoning steps (i.e., thoughts) in CoT traces, enabling supervised model training on length-reduced CoT data while preserving both answer accuracy and the model's ability to generate coherent reasoning. Our rationale is that CoT traces typically follow a three-stage structure: problem understanding, exploration, and solution convergence. Through empirical analysis, we find that retaining the structure of the reasoning trace, especially the early stage of problem understanding (rich in reflective cues) and the final stage of solution convergence, is sufficient to achieve lossless reasoning supervision. To this end, we propose an Edge-Preserving Condensation method, EPiC, which selectively retains only the initial and final segments of each CoT trace while discarding the middle portion. This design draws an analogy to preserving the "edge" of a reasoning trajectory, capturing both the initial problem framing and the final answer synthesis, to maintain logical continuity. Experiments across multiple model families (Qwen and LLaMA) and benchmarks show that EPiC reduces training time by over 34% while achieving lossless reasoning accuracy on MATH500, comparable to full CoT supervision. To the best of our knowledge, this is the first study to explore thought-level CoT condensation for efficient reasoning model distillation.

full paperreport(pdf)

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

full paperreport(pdf)

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s per NPU even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

full paperreport(pdf)

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration such as evaluation batch size, GPU count, and GPU version can introduce significant difference in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision -- while critical for reproducibility -- is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

full paperreport(pdf)

Embedding-based search is widely used in applications such as recommendation and retrieval-augmented generation (RAG). Recently, there is a growing demand to support these capabilities over personal data stored locally on devices. However, maintaining the necessary data structure associated with the embedding-based search is often infeasible due to its high storage overhead. For example, indexing 100 GB of raw data requires 150 to 700 GB of storage, making local deployment impractical. Reducing this overhead while maintaining search quality and latency becomes a critical challenge. In this paper, we present LEANN, a storage-efficient approximate nearest neighbor (ANN) search index optimized for resource-constrained personal devices. LEANN combines a compact graph-based structure with an efficient on-the-fly recomputation strategy to enable fast and accurate retrieval with minimal storage overhead. Our evaluation shows that LEANN reduces index size to under 5% of the original raw data, achieving up to 50 times smaller storage than standard indexes, while maintaining 90% top-3 recall in under 2 seconds on real-world question answering benchmarks.

full paperreport(pdf)

This paper introduces EdgeProfiler, a fast profiling framework designed for evaluating lightweight Large Language Models (LLMs) on edge systems. While LLMs offer remarkable capabilities in natural language understanding and generation, their high computational, memory, and power requirements often confine them to cloud environments. EdgeProfiler addresses these challenges by providing a systematic methodology for assessing LLM performance in resource-constrained edge settings. The framework profiles compact LLMs, including TinyLLaMA, Gemma3.1B, Llama3.2-1B, and DeepSeek-r1-1.5B, using aggressive quantization techniques and strict memory constraints. Analytical modeling is used to estimate latency, FLOPs, and energy consumption. The profiling reveals that 4-bit quantization reduces model memory usage by approximately 60-70%, while maintaining accuracy within 2-5% of full-precision baselines. Inference speeds are observed to improve by 2-3x compared to FP16 baselines across various edge devices. Power modeling estimates a 35-50% reduction in energy consumption for INT4 configurations, enabling practical deployment on hardware such as Raspberry Pi 4/5 and Jetson Orin Nano Super. Our findings emphasize the importance of efficient profiling tailored to lightweight LLMs in edge environments, balancing accuracy, energy efficiency, and computational feasibility.

full paperreport(pdf)

This paper introduces EdgeProfiler, a fast profiling framework designed for evaluating lightweight Large Language Models (LLMs) on edge systems. While LLMs offer remarkable capabilities in natural language understanding and generation, their high computational, memory, and power requirements often confine them to cloud environments. EdgeProfiler addresses these challenges by providing a systematic methodology for assessing LLM performance in resource-constrained edge settings. The framework profiles compact LLMs, including TinyLLaMA, Gemma3.1B, Llama3.2-1B, and DeepSeek-r1-1.5B, using aggressive quantization techniques and strict memory constraints. Analytical modeling is used to estimate latency, FLOPs, and energy consumption. The profiling reveals that 4-bit quantization reduces model memory usage by approximately 60-70%, while maintaining accuracy within 2-5% of full-precision baselines. Inference speeds are observed to improve by 2-3x compared to FP16 baselines across various edge devices. Power modeling estimates a 35-50% reduction in energy consumption for INT4 configurations, enabling practical deployment on hardware such as Raspberry Pi 4/5 and Jetson Orin Nano Super. Our findings emphasize the importance of efficient profiling tailored to lightweight LLMs in edge environments, balancing accuracy, energy efficiency, and computational feasibility.

full paperreport(pdf)

Large language models (LLMs) have shown remarkable reasoning capabilities when trained with chain-of-thought (CoT) supervision. However, the long and verbose CoT traces, especially those distilled from large reasoning models (LRMs) such as DeepSeek-R1, significantly increase training costs during the distillation process, where a non-reasoning base model is taught to replicate the reasoning behavior of an LRM. In this work, we study the problem of CoT condensation for resource-efficient reasoning training, aimed at pruning intermediate reasoning steps (i.e., thoughts) in CoT traces, enabling supervised model training on length-reduced CoT data while preserving both answer accuracy and the model's ability to generate coherent reasoning. Our rationale is that CoT traces typically follow a three-stage structure: problem understanding, exploration, and solution convergence. Through empirical analysis, we find that retaining the structure of the reasoning trace, especially the early stage of problem understanding (rich in reflective cues) and the final stage of solution convergence, is sufficient to achieve lossless reasoning supervision. To this end, we propose an Edge-Preserving Condensation method, EPiC, which selectively retains only the initial and final segments of each CoT trace while discarding the middle portion. This design draws an analogy to preserving the "edge" of a reasoning trajectory, capturing both the initial problem framing and the final answer synthesis, to maintain logical continuity. Experiments across multiple model families (Qwen and LLaMA) and benchmarks show that EPiC reduces training time by over 34% while achieving lossless reasoning accuracy on MATH500, comparable to full CoT supervision. To the best of our knowledge, this is the first study to explore thought-level CoT condensation for efficient reasoning model distillation.

full paperreport(pdf)

Large Language Models (LLMs) have recently achieved remarkable progress by leveraging Reinforcement Learning and extended Chain-of-Thought (CoT) techniques. However, the challenge of performing efficient language reasoning--especially during inference with extremely long outputs--has drawn increasing attention from the research community. In this work, we propose a dynamic ratio-based training pipeline that does not rely on sophisticated data annotations or interpolation between multiple models. We continuously balance the weights between the model's System-1 and System-2 data to eliminate redundant reasoning processes while preserving the model's reasoning capability. We validate our approach across models on DeepSeek-R1-Distill-7B and DeepSeek-R1-Distill-14B and on a diverse set of benchmarks with varying difficulty levels. Our method significantly reduces the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning. Our code and data will be available soon.

full paperreport(pdf)

While large language models (LLMs) have shown promising capabilities as zero-shot planners for embodied agents, their inability to learn from experience and build persistent mental models limits their robustness in complex open-world environments like Minecraft. We introduce MINDSTORES, an experience-augmented planning framework that enables embodied agents to build and leverage mental models through natural interaction with their environment. Drawing inspiration from how humans construct and refine cognitive mental models, our approach extends existing zero-shot LLM planning by maintaining a database of past experiences that informs future planning iterations. The key innovation is representing accumulated experiences as natural language embeddings of (state, task, plan, outcome) tuples, which can then be efficiently retrieved and reasoned over by an LLM planner to generate insights and guide plan refinement for novel states and tasks. Through extensive experiments in the MineDojo environment, a simulation environment for agents in Minecraft that provides low-level controls for Minecraft, we find that MINDSTORES learns and applies its knowledge significantly better than existing memory-based LLM planners while maintaining the flexibility and generalization benefits of zero-shot approaches, representing an important step toward more capable embodied AI systems that can learn continuously through natural experience.

full paperreport(pdf)

Vision-and-Language Navigation (VLN) presents a complex challenge in embodied AI, requiring agents to interpret natural language instructions and navigate through visually rich, unfamiliar environments. Recent advances in large vision-language models (LVLMs), such as CLIP and Flamingo, have significantly improved multimodal understanding but introduced new challenges related to computational cost and real-time deployment. In this project, we propose a modular, plug-and-play navigation framework that decouples vision-language understanding from action planning. By integrating a frozen vision-language model, Qwen2.5-VL-7B-Instruct, with lightweight planning logic, we aim to achieve flexible, fast, and adaptable navigation without extensive model fine-tuning. Our framework leverages prompt engineering, structured history management, and a two-frame visual input strategy to enhance decision-making continuity across navigation steps. We evaluate our system on the Room-to-Room benchmark within the VLN-CE setting using the Matterport3D dataset and Habitat-Lab simulation environment. Although our initial results reveal challenges in generalizing to unseen environments under strict evaluation settings, our modular approach lays a foundation for scalable and efficient navigation systems, highlighting promising directions for future improvement through enhanced environmental priors and expanded multimodal input integration.

full paperreport(pdf)

Medical vision-language alignment through cross-modal contrastive learning shows promising performance in image-text matching tasks, such as retrieval and zero-shot classification. However, conventional cross-modal contrastive learning (CLIP-based) methods suffer from suboptimal visual representation capabilities, which also limits their effectiveness in vision-language alignment. In contrast, although the models pretrained via multimodal masked modeling struggle with direct cross-modal matching, they excel in visual representation. To address this contradiction, we propose ALTA (ALign Through Adapting), an efficient medical vision-language alignment method that utilizes only about 8% of the trainable parameters and less than 1/5 of the computational consumption required for masked record modeling. ALTA achieves superior performance in vision-language matching tasks like retrieval and zero-shot classification by adapting the pretrained vision model from masked record modeling. Additionally, we integrate temporal-multiview radiograph inputs to enhance the information consistency between radiographs and their corresponding descriptions in reports, further improving the vision-language alignment. Experimental evaluations show that ALTA outperforms the best-performing counterpart by over 4% absolute points in text-to-image accuracy and approximately 6% absolute points in image-to-text retrieval accuracy. The adaptation of vision-language models during efficient alignment also promotes better vision and language understanding. Code is publicly available at https://github.com/DopamineLcy/ALTA.

full paperreport(pdf)

Recent advancements in multimodal large language models have successfully extended the Reason-Then-Respond paradigm to image-based reasoning, yet video-based reasoning remains an underdeveloped frontier, primarily due to the scarcity of high-quality reasoning-oriented data and effective training methodologies. To bridge this gap, we introduce DarkEventInfer and MixVidQA, two novel datasets specifically designed to stimulate the model's advanced video understanding and reasoning abilities. DarkEventinfer presents videos with masked event segments, requiring models to infer the obscured content based on contextual video cues. MixVidQA, on the other hand, presents interleaved video sequences composed of two distinct clips, challenging models to isolate and reason about one while disregarding the other. Leveraging these carefully curated training samples together with reinforcement learning guided by diverse reward functions, we develop VersaVid-R1, the first versatile video understanding and reasoning model under the Reason-Then-Respond paradigm capable of handling multiple-choice and open-ended question answering, as well as video captioning tasks. Extensive experiments demonstrate that VersaVid-R1 significantly outperforms existing models across a broad spectrum of benchmarks, covering video general understanding, cognitive reasoning, and captioning tasks.

full paperreport(pdf)

Large multimodal models (LMMs) have shown remarkable progress in audio-visual understanding, yet they struggle with real-world scenarios that require complex reasoning across extensive video collections. Existing benchmarks for video question answering remain limited in scope, typically involving one clip per query, which falls short of representing the challenges of large-scale, audio-visual retrieval and reasoning encountered in practical applications. To bridge this gap, we introduce a novel task named AV-HaystacksQA, where the goal is to identify salient segments across different videos in response to a query and link them together to generate the most informative answer. To this end, we present AVHaystacks, an audio-visual benchmark comprising 3100 annotated QA pairs designed to assess the capabilities of LMMs in multi-video retrieval and temporal grounding task. Additionally, we propose a model-agnostic, multi-agent framework MAGNET to address this challenge, achieving up to 89% and 65% relative improvements over baseline methods on BLEU@4 and GPT evaluation scores in QA task on our proposed AVHaystacks. To enable robust evaluation of multi-video retrieval and temporal grounding for optimal response generation, we introduce two new metrics, STEM, which captures alignment errors between a ground truth and a predicted step sequence and MTGS, to facilitate balanced and interpretable evaluation of segment-level grounding performance. Project: https://schowdhury671.github.io/magnet_project/

full paperreport(pdf)

Despite Contrastive Language-Image Pretraining (CLIP)'s remarkable capability to retrieve content across modalities, a substantial modality gap persists in its feature space. Intriguingly, we discover that off-the-shelf MLLMs (Multimodal Large Language Models) demonstrate powerful inherent modality alignment properties. While recent MLLM-based retrievers with unified architectures partially mitigate this gap, their reliance on coarse modality alignment mechanisms fundamentally limits their potential. In this work, We introduce MAPLE (Modality-Aligned Preference Learning for Embeddings), a novel framework that leverages the fine grained alignment priors inherent in MLLM to guide cross modal representation learning. MAPLE formulates the learning process as reinforcement learning with two key components: (1) Automatic preference data construction using off-the-shelf MLLM, and (2) a new Relative Preference Alignment (RPA) loss, which adapts Direct Preference Optimization (DPO) to the embedding learning setting. Experimental results show that our preference-guided alignment achieves substantial gains in fine-grained cross-modal retrieval, underscoring its effectiveness in handling nuanced semantic distinctions.

full paperreport(pdf)

Large multimodal models (LMMs) have shown remarkable progress in audio-visual understanding, yet they struggle with real-world scenarios that require complex reasoning across extensive video collections. Existing benchmarks for video question answering remain limited in scope, typically involving one clip per query, which falls short of representing the challenges of large-scale, audio-visual retrieval and reasoning encountered in practical applications. To bridge this gap, we introduce a novel task named AV-HaystacksQA, where the goal is to identify salient segments across different videos in response to a query and link them together to generate the most informative answer. To this end, we present AVHaystacks, an audio-visual benchmark comprising 3100 annotated QA pairs designed to assess the capabilities of LMMs in multi-video retrieval and temporal grounding task. Additionally, we propose a model-agnostic, multi-agent framework MAGNET to address this challenge, achieving up to 89% and 65% relative improvements over baseline methods on BLEU@4 and GPT evaluation scores in QA task on our proposed AVHaystacks. To enable robust evaluation of multi-video retrieval and temporal grounding for optimal response generation, we introduce two new metrics, STEM, which captures alignment errors between a ground truth and a predicted step sequence and MTGS, to facilitate balanced and interpretable evaluation of segment-level grounding performance. Project: https://schowdhury671.github.io/magnet_project/

full paperreport(pdf)

Advances in multimodal text-image models have enabled effective text-based querying in extensive image collections. While these models show convincing performance for everyday life scenes, querying in highly homogeneous, specialized domains remains challenging. The primary problem is that users can often provide only vague textual descriptions as they lack expert knowledge to discriminate between homogenous entities. This work investigates whether adding location-based prompts to complement these vague text queries can enhance retrieval performance. Specifically, we collected a dataset of 741 human annotations, each containing short and long textual descriptions and bounding boxes indicating regions of interest in challenging underwater scenes. Using these annotations, we evaluate the performance of CLIP when queried on various static sub-regions of images compared to the full image. Our results show that both a simple 3-by-3 partitioning and a 5-grid overlap significantly improve retrieval effectiveness and remain robust to perturbations of the annotation box.

full paperreport(pdf)

Composed image retrieval (CIR) allows a user to locate a target image by applying a fine-grained textual edit (e.g., ``turn the dress blue'' or ``remove stripes'') to a reference image. Zero-shot CIR, which embeds the image and the text with separate pretrained vision-language encoders, reaches only 20-25\% Recall@10 on the FashionIQ benchmark. We improve this by fine-tuning BLIP-2 with a lightweight Q-Former that fuses visual and textual features into a single embedding, raising Recall@10 to 45.6\% (shirt), 40.1\% (dress), and 50.4\% (top-tee) and increasing the average Recall@50 to 67.6\%. We also examine Retrieval-DPO, which fine-tunes CLIP's text encoder with a Direct Preference Optimization loss applied to FAISS-mined hard negatives. Despite extensive tuning of the scaling factor, index, and sampling strategy, Retrieval-DPO attains only 0.02\% Recall@10 -- far below zero-shot and prompt-tuned baselines -- because it (i) lacks joint image-text fusion, (ii) uses a margin objective misaligned with top-$K$ metrics, (iii) relies on low-quality negatives, and (iv) keeps the vision and Transformer layers frozen. Our results show that effective preference-based CIR requires genuine multimodal fusion, ranking-aware objectives, and carefully curated negatives.

full paperreport(pdf)

Composed Video Retrieval (CoVR) retrieves a target video given a query video and a modification text describing the intended change. Existing CoVR benchmarks emphasize appearance shifts or coarse event changes and therefore do not test the ability to capture subtle, fast-paced temporal differences. We introduce TF-CoVR, the first large-scale benchmark dedicated to temporally fine-grained CoVR. TF-CoVR focuses on gymnastics and diving and provides 180K triplets drawn from FineGym and FineDiving. Previous CoVR benchmarks focusing on temporal aspect, link each query to a single target segment taken from the same video, limiting practical usefulness. In TF-CoVR, we instead construct each <query, modification> pair by prompting an LLM with the label differences between clips drawn from different videos; every pair is thus associated with multiple valid target videos (3.9 on average), reflecting real-world tasks such as sports-highlight generation. To model these temporal dynamics we propose TF-CoVR-Base, a concise two-stage training framework: (i) pre-train a video encoder on fine-grained action classification to obtain temporally discriminative embeddings; (ii) align the composed query with candidate videos using contrastive learning. We conduct the first comprehensive study of image, video, and general multimodal embedding (GME) models on temporally fine-grained composed retrieval in both zero-shot and fine-tuning regimes. On TF-CoVR, TF-CoVR-Base improves zero-shot mAP@50 from 5.92 (LanguageBind) to 7.51, and after fine-tuning raises the state-of-the-art from 19.83 to 25.82.

full paperreport(pdf)