Daily Papers

Different retrievers exist for retrieval-augmented generation, yet none performs best for all queries. Therefore, the key question is: given a query and retrieved passages from different retrievers, which one is most suited for that query? Existing routers usually make predictions from query-specific features and a fixed pool of retrievers; such routers have no output for an unseen retriever and are incapable by design on a new corpus. To fill these gaps, RAGCompass reads the retrieved passage sets themselves: passages from an unfamiliar retriever or corpus stay directly comparable to other passages, so routing stays well designed where a query-only map would fail. We contrastively fine-tune a reranker against the answer each evidence set produces, which accounts for most of the gain; a small comparison head adds a moderate improvement, and neither component has retriever-specific parameters. RAGCompass improves answer F1 by 28% relative to the best fixed retriever, and outperforms fusion baselines, three query-only routers, and a prompted 8B language model. Generalization experiments show that RAGCompass keeps most of its gain under an unseen corpus, retriever, or reader, while query-only baselines fall to the best fixed retriever or below.

Reinforcement learning from human feedback (RLHF) is effective at aligning large language models (LLMs) to human preferences, but gathering high quality human preference labels is a key bottleneck. We conduct a head-to-head comparison of RLHF vs. RL from AI Feedback (RLAIF) - a technique where preferences are labeled by an off-the-shelf LLM in lieu of humans, and we find that they result in similar improvements. On the task of summarization, human evaluators prefer generations from both RLAIF and RLHF over a baseline supervised fine-tuned model in ~70% of cases. Furthermore, when asked to rate RLAIF vs. RLHF summaries, humans prefer both at equal rates. These results suggest that RLAIF can yield human-level performance, offering a potential solution to the scalability limitations of RLHF.

The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy. We introduce Megalodon, a neural architecture for efficient sequence modeling with unlimited context length. Megalodon inherits the architecture of Mega (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability and stability, including complex exponential moving average (CEMA), timestep normalization layer, normalized attention mechanism and pre-norm with two-hop residual configuration. In a controlled head-to-head comparison with Llama2, Megalodon achieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens. Megalodon reaches a training loss of 1.70, landing mid-way between Llama2-7B (1.75) and 13B (1.67). Code: https://github.com/XuezheMax/megalodon

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Positron emission tomography (PET) is a cornerstone of modern oncologic and neurologic imaging, distinguished by its unique ability to illuminate dynamic metabolic processes that transcend the anatomical focus of traditional imaging technologies. Radiology reports are essential for clinical decision making, yet their manual creation is labor-intensive and time-consuming. Recent advancements of vision-language models (VLMs) have shown strong potential in medical applications, presenting a promising avenue for automating report generation. However, existing applications of VLMs in the medical domain have predominantly focused on structural imaging modalities, while the unique characteristics of molecular PET imaging have largely been overlooked. To bridge the gap, we introduce PET2Rep, a large-scale comprehensive benchmark for evaluation of general and medical VLMs for radiology report generation for PET images. PET2Rep stands out as the first dedicated dataset for PET report generation with metabolic information, uniquely capturing whole-body image-report pairs that cover dozens of organs to fill the critical gap in existing benchmarks and mirror real-world clinical comprehensiveness. In addition to widely recognized natural language generation metrics, we introduce a series of clinical efficiency metrics to evaluate the quality of radiotracer uptake pattern description in key organs in generated reports. We conduct a head-to-head comparison of 30 cutting-edge general-purpose and medical-specialized VLMs. The results show that the current state-of-the-art VLMs perform poorly on PET report generation task, falling considerably short of fulfilling practical needs. Moreover, we identify several key insufficiency that need to be addressed to advance the development in medical applications.

Multimodal large language models are increasingly deployed as long-horizon agents, where memory must do more than recall: it must track an evolving world, revise what has gone stale, and surface the right evidence at decision time. Existing benchmarks measure recall over static dialogue, collapse memory into a single end-of-task accuracy, and reduce visual observations to captions, leaving us unable to localize failures to writing, maintenance, retrieval, or use. The rise of agent harnesses that author their own memory sharpens this gap, since we have no principled way to compare hand-designed pipelines with self-managing alternatives. To close these gaps, we formulate multimodal agent memory as an Action-World Interaction Loop with an observable four-stage lifecycle, and instantiate it in WorldMemArena: 400 multi-session multimodal tasks spanning Lifelong Evolution (evolving personal and task states) and Agentic Execution (memory from real observations, actions, and feedback), annotated with gold memory points, updates, distractors, and evidence chains for stage-level diagnosis. This enables the first head-to-head comparison of long-context, manually designed (RAG and external memory systems), and harness-based memory agents. Results show that: (1) better memory writing and storage do not guarantee better performance; (2) multimodal memory still struggles to fully use visual evidence; (3) systems are unstable across domains and degrade on realistic agentic trajectories; and (4) harness memory is more flexible but remains costly and less reliable.

Given the impact of language models on the field of Natural Language Processing, a number of Spanish encoder-only masked language models (aka BERTs) have been trained and released. These models were developed either within large projects using very large private corpora or by means of smaller scale academic efforts leveraging freely available data. In this paper we present a comprehensive head-to-head comparison of language models for Spanish with the following results: (i) Previously ignored multilingual models from large companies fare better than monolingual models, substantially changing the evaluation landscape of language models in Spanish; (ii) Results across the monolingual models are not conclusive, with supposedly smaller and inferior models performing competitively. Based on these empirical results, we argue for the need of more research to understand the factors underlying them. In this sense, the effect of corpus size, quality and pre-training techniques need to be further investigated to be able to obtain Spanish monolingual models significantly better than the multilingual ones released by large private companies, specially in the face of rapid ongoing progress in the field. The recent activity in the development of language technology for Spanish is to be welcomed, but our results show that building language models remains an open, resource-heavy problem which requires to marry resources (monetary and/or computational) with the best research expertise and practice.

Large language models have made significant progress in complex but easy-to-verify problems, yet they still struggle with discovering the unknown. In this paper, we present AlphaResearch, an autonomous research agent designed to discover new algorithms on open-ended problems. To synergize the feasibility and innovation of the discovery process, we construct a novel dual research environment by combining the execution-based verify and simulated real-world peer review environment. AlphaResearch discovers new algorithm by iteratively running the following steps: (1) propose new ideas (2) verify the ideas in the dual research environment (3) optimize the research proposals for better performance. To promote a transparent evaluation process, we construct AlphaResearchComp, a new evaluation benchmark that includes an eight open-ended algorithmic problems competition, with each problem carefully curated and verified through executable pipelines, objective metrics, and reproducibility checks. AlphaResearch gets a 2/8 win rate in head-to-head comparison with human researchers, demonstrate the possibility of accelerating algorithm discovery with LLMs. Notably, the algorithm discovered by AlphaResearch on the ``packing circles'' problem achieves the best-of-known performance, surpassing the results of human researchers and strong baselines from recent work (e.g., AlphaEvolve). Additionally, we conduct a comprehensive analysis of the remaining challenges of the 6/8 failure cases, providing valuable insights for future research.

Text embeddings for Portuguese have no dedicated benchmark: evaluation rests on translated corpora such as English MS MARCO or on thin multilingual coverage, with native tasks scattered and unconsolidated. We introduce MTEB-PT, a benchmark of 22 native Brazilian-Portuguese tasks across seven categories (classification, multilabel classification, pair classification, semantic textual similarity, clustering, retrieval, and reranking), admitting only data created or found in Portuguese and excluding translations by construction. We evaluate 93 models spanning 23M to 27B parameters: 73 open-weight and 20 closed commercial APIs. Alongside the leaderboard we report a statistical layer for every headline comparison: per-task bootstrap confidence intervals, paired-bootstrap significance, a task- and instance-level discrimination analysis (how sharply each task separates models) adapted from Item Response Theory, and a cross-leaderboard correlation. Three findings stand out. The benchmark cleanly separates about a dozen tiers of models, though the top six are statistically too close to order. An openly licensed, self-hostable model reaches that leading tier, so strong Portuguese embedding quality does not require a commercial API. And a model's rank on the global multilingual leaderboard predicts its Portuguese rank only moderately (Spearman rho = 0.75 over 55 shared models; one model ranks 3rd there and 49th here), so a native benchmark measures something the multilingual boards do not. We release every task, our code, and a public leaderboard, so practitioners can choose Portuguese embedding models on native evidence.

The advent of scalable deep models and large datasets has improved the performance of Neural Machine Translation. Knowledge Distillation (KD) enhances efficiency by transferring knowledge from a teacher model to a more compact student model. However, KD approaches to Transformer architecture often rely on heuristics, particularly when deciding which teacher layers to distill from. In this paper, we introduce the 'Align-to-Distill' (A2D) strategy, designed to address the feature mapping problem by adaptively aligning student attention heads with their teacher counterparts during training. The Attention Alignment Module in A2D performs a dense head-by-head comparison between student and teacher attention heads across layers, turning the combinatorial mapping heuristics into a learning problem. Our experiments show the efficacy of A2D, demonstrating gains of up to +3.61 and +0.63 BLEU points for WMT-2022 De->Dsb and WMT-2014 En->De, respectively, compared to Transformer baselines.

In Natural Language Processing (NLP), the Elo rating system, originally designed for ranking players in dynamic games such as chess, is increasingly being used to evaluate Large Language Models (LLMs) through "A vs B" paired comparisons. However, while popular, the system's suitability for assessing entities with constant skill levels, such as LLMs, remains relatively unexplored. We study two fundamental axioms that evaluation methods should adhere to: reliability and transitivity. We conduct extensive evaluation of Elo behaviour, illustrating that individual Elo computations exhibit volatility and delving into the impact of varying the Elo rating system's hyperparameters. We show that these axioms are not always satisfied raising questions about the reliability of current comparative evaluations of LLMs. If the current use of Elo scores is intended to substitute the costly head-to-head comparison of LLMs, it is crucial to ensure the ranking is as robust as possible. Guided by the axioms, our findings offer concrete guidelines for enhancing the reliability of LLM evaluation methods, suggesting a need for reassessment of existing comparative approaches.

Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. We also demonstrate the effectiveness of this reward model at aligning models to follow instructions in RLHF. We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2 and openly release the trained Reward Model at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward

A knowledge gap persists between machine learning (ML) developers (e.g., data scientists) and practitioners (e.g., clinicians), hampering the full utilization of ML for clinical data analysis. We investigated the potential of the ChatGPT Advanced Data Analysis (ADA), an extension of GPT-4, to bridge this gap and perform ML analyses efficiently. Real-world clinical datasets and study details from large trials across various medical specialties were presented to ChatGPT ADA without specific guidance. ChatGPT ADA autonomously developed state-of-the-art ML models based on the original study's training data to predict clinical outcomes such as cancer development, cancer progression, disease complications, or biomarkers such as pathogenic gene sequences. Following the re-implementation and optimization of the published models, the head-to-head comparison of the ChatGPT ADA-crafted ML models and their respective manually crafted counterparts revealed no significant differences in traditional performance metrics (P>0.071). Strikingly, the ChatGPT ADA-crafted ML models often outperformed their counterparts. In conclusion, ChatGPT ADA offers a promising avenue to democratize ML in medicine by simplifying complex data analyses, yet should enhance, not replace, specialized training and resources, to promote broader applications in medical research and practice.

Recent advancements in large language models (LLMs) have sparked optimism about their potential to accelerate scientific discovery, with a growing number of works proposing research agents that autonomously generate and validate new ideas. Despite this, no evaluations have shown that LLM systems can take the very first step of producing novel, expert-level ideas, let alone perform the entire research process. We address this by establishing an experimental design that evaluates research idea generation while controlling for confounders and performs the first head-to-head comparison between expert NLP researchers and an LLM ideation agent. By recruiting over 100 NLP researchers to write novel ideas and blind reviews of both LLM and human ideas, we obtain the first statistically significant conclusion on current LLM capabilities for research ideation: we find LLM-generated ideas are judged as more novel (p < 0.05) than human expert ideas while being judged slightly weaker on feasibility. Studying our agent baselines closely, we identify open problems in building and evaluating research agents, including failures of LLM self-evaluation and their lack of diversity in generation. Finally, we acknowledge that human judgements of novelty can be difficult, even by experts, and propose an end-to-end study design which recruits researchers to execute these ideas into full projects, enabling us to study whether these novelty and feasibility judgements result in meaningful differences in research outcome.

Detecting abnormal events in video is commonly framed as a one-class classification task, where training videos contain only normal events, while test videos encompass both normal and abnormal events. In this scenario, anomaly detection is an open-set problem. However, some studies assimilate anomaly detection to action recognition. This is a closed-set scenario that fails to test the capability of systems at detecting new anomaly types. To this end, we propose UBnormal, a new supervised open-set benchmark composed of multiple virtual scenes for video anomaly detection. Unlike existing data sets, we introduce abnormal events annotated at the pixel level at training time, for the first time enabling the use of fully-supervised learning methods for abnormal event detection. To preserve the typical open-set formulation, we make sure to include disjoint sets of anomaly types in our training and test collections of videos. To our knowledge, UBnormal is the first video anomaly detection benchmark to allow a fair head-to-head comparison between one-class open-set models and supervised closed-set models, as shown in our experiments. Moreover, we provide empirical evidence showing that UBnormal can enhance the performance of a state-of-the-art anomaly detection framework on two prominent data sets, Avenue and ShanghaiTech. Our benchmark is freely available at https://github.com/lilygeorgescu/UBnormal.

Large language models are increasingly capable of generating fluent-appearing text with relatively little task-specific supervision. But can these models accurately explain classification decisions? We consider the task of generating free-text explanations using human-written examples in a few-shot manner. We find that (1) authoring higher quality prompts results in higher quality generations; and (2) surprisingly, in a head-to-head comparison, crowdworkers often prefer explanations generated by GPT-3 to crowdsourced explanations in existing datasets. Our human studies also show, however, that while models often produce factual, grammatical, and sufficient explanations, they have room to improve along axes such as providing novel information and supporting the label. We create a pipeline that combines GPT-3 with a supervised filter that incorporates binary acceptability judgments from humans in the loop. Despite the intrinsic subjectivity of acceptability judgments, we demonstrate that acceptability is partially correlated with various fine-grained attributes of explanations. Our approach is able to consistently filter GPT-3-generated explanations deemed acceptable by humans.

Medical Vision-Language Pretraining (MedVLP) shows promise in learning generalizable and transferable visual representations from paired and unpaired medical images and reports. MedVLP can provide useful features to downstream tasks and facilitate adapting task-specific models to new setups using fewer examples. However, existing MedVLP methods often differ in terms of datasets, preprocessing, and finetuning implementations. This pose great challenges in evaluating how well a MedVLP method generalizes to various clinically-relevant tasks due to the lack of unified, standardized, and comprehensive benchmark. To fill this gap, we propose BenchX, a unified benchmark framework that enables head-to-head comparison and systematical analysis between MedVLP methods using public chest X-ray datasets. Specifically, BenchX is composed of three components: 1) Comprehensive datasets covering nine datasets and four medical tasks; 2) Benchmark suites to standardize data preprocessing, train-test splits, and parameter selection; 3) Unified finetuning protocols that accommodate heterogeneous MedVLP methods for consistent task adaptation in classification, segmentation, and report generation, respectively. Utilizing BenchX, we establish baselines for nine state-of-the-art MedVLP methods and found that the performance of some early MedVLP methods can be enhanced to surpass more recent ones, prompting a revisiting of the developments and conclusions from prior works in MedVLP. Our code are available at https://github.com/yangzhou12/BenchX.

Vision-Language-Action (VLA) models have recently been proposed as a pathway toward generalist robotic policies capable of interpreting natural language and visual inputs to generate manipulation actions. However, their effectiveness and efficiency on structured, long-horizon manipulation tasks remain unclear. In this work, we present a head-to-head empirical comparison between a fine-tuned open-weight VLA model π0 and a neuro-symbolic architecture that combines PDDL-based symbolic planning with learned low-level control. We evaluate both approaches on structured variants of the Towers of Hanoi manipulation task in simulation while measuring both task performance and energy consumption during training and execution. On the 3-block task, the neuro-symbolic model achieves 95% success compared to 34% for the best-performing VLA. The neuro-symbolic model also generalizes to an unseen 4-block variant (78% success), whereas both VLAs fail to complete the task. During training, VLA fine-tuning consumes nearly two orders of magnitude more energy than the neuro-symbolic approach. These results highlight important trade-offs between end-to-end foundation-model approaches and structured reasoning architectures for long-horizon robotic manipulation, emphasizing the role of explicit symbolic structure in improving reliability, data efficiency, and energy efficiency. Code and models are available at https://price-is-not-right.github.io

Large-scale diffusion models have achieved state-of-the-art results on text-to-image synthesis (T2I) tasks. Despite their ability to generate high-quality yet creative images, we observe that attribution-binding and compositional capabilities are still considered major challenging issues, especially when involving multiple objects. In this work, we improve the compositional skills of T2I models, specifically more accurate attribute binding and better image compositions. To do this, we incorporate linguistic structures with the diffusion guidance process based on the controllable properties of manipulating cross-attention layers in diffusion-based T2I models. We observe that keys and values in cross-attention layers have strong semantic meanings associated with object layouts and content. Therefore, we can better preserve the compositional semantics in the generated image by manipulating the cross-attention representations based on linguistic insights. Built upon Stable Diffusion, a SOTA T2I model, our structured cross-attention design is efficient that requires no additional training samples. We achieve better compositional skills in qualitative and quantitative results, leading to a 5-8% advantage in head-to-head user comparison studies. Lastly, we conduct an in-depth analysis to reveal potential causes of incorrect image compositions and justify the properties of cross-attention layers in the generation process.

We describe an analysis comparing the pp elastic cross section as measured by the D0 Collaboration at a center-of-mass energy of 1.96 TeV to that in pp collisions as measured by the TOTEM Collaboration at 2.76, 7, 8, and 13 TeV using a model-independent approach. The TOTEM cross sections extrapolated to a center-of-mass energy of s = 1.96 TeV are compared with the D0 measurement in the region of the diffractive minimum and the second maximum of the pp cross section. The two data sets disagree at the 3.4σ level and thus provide evidence for the t-channel exchange of a colorless, C-odd gluonic compound, also known as the odderon. We combine these results with a TOTEM analysis of the same C-odd exchange based on the total cross section and the ratio of the real to imaginary parts of the forward elastic scattering amplitude in pp scattering. The combined significance of these results is larger than 5σ and is interpreted as the first observation of the exchange of a colorless, C-odd gluonic compound.

In this paper we introduce a new synchronisation task, Gesture-Sync: determining if a person's gestures are correlated with their speech or not. In comparison to Lip-Sync, Gesture-Sync is far more challenging as there is a far looser relationship between the voice and body movement than there is between voice and lip motion. We introduce a dual-encoder model for this task, and compare a number of input representations including RGB frames, keypoint images, and keypoint vectors, assessing their performance and advantages. We show that the model can be trained using self-supervised learning alone, and evaluate its performance on the LRS3 dataset. Finally, we demonstrate applications of Gesture-Sync for audio-visual synchronisation, and in determining who is the speaker in a crowd, without seeing their faces. The code, datasets and pre-trained models can be found at: https://www.robots.ox.ac.uk/~vgg/research/gestsync.

Adaptive computation aims to improve language-model inference by tailoring execution to each input. For layer programs, oracle evaluations use known answers to estimate the potential gain from this flexibility, before a practical selector is available. However, a gain from selection does not by itself explain why the chosen programs help. This study examines this distinction using 32 layer-skipping and repetition programs on two models and 4,413 multiple-choice items. The analysis compares their gains over a fixed action selected without the evaluation prompt with those of input-blind perturbations at the same sites, re-evaluating selections on another prompt. With shared option order, the controls give 10.2-11.8 and 15.6-19.4 percentage points of headroom on Qwen3-4B-Base and Llama-3.1-8B, exceeding the real programs' 9.0 and 10.1 in all three random-direction draws per model. They match answer-change rate only, and the ordering depends on the menu: in post hoc comparisons, real programs lead on Llama's repeat-only menu in every draw. A smaller KL-calibrated comparison, including an input-dependent control, favours real programs in point estimate, with inconclusive corrected tests. Fixed letter offsets produce headroom of similar scale. Rotating options sharply reduces both families' headroom, while leaving positive real-minus-control differences of 1.4-2.3 and 3.7-4.5 points; their magnitudes and statistical support depend on further adjustments and the reference. A supplementary generated-answer test finds that search-selected programs keep a 26.0-point advantage over programs selected for other problems after rewording, without a placebo comparison. These results show that substantial headroom can persist across prompts with shared option order without establishing a benefit specific to the selected layer computation; neither ordering against these controls identifies that benefit.

The design of attention mechanisms in long-context large language models faces an ``impossible triangle'' among KV cache size, training and prefill FLOPs, and model quality. GQA achieves strong quality but incurs high KV cache overhead; alternative approaches such as MLA and MFA can compress the cache, but MLA suffers from quality loss while MFA substantially increases training and prefill compute. We propose Virtual Head Attention (VHA), which simultaneously improves all three vertices through two lightweight linear operations: Q Premix applies a near-identity transformation to queries within each KV group to recover virtual head diversity from halved physical query heads, and Linear Postmix fuses inter-head features via a low-rank I + AB^top residual structure. Since VHA introduces no nonlinear operations, Premix can be folded into the query projection and Postmix into the output projection at inference time, reducing the model to a standard GQA-2 architecture that directly reuses all existing GQA inference optimizations. With only two KV heads, the group size in large models readily reaches 64, fully exploiting Tensor Core parallelism during decoding. Across four scales with 4K-context 50B-token pretraining, VHA consistently surpasses the standard GQA baseline at every scale (+0.05\% to +0.81\%), while reducing KV cache by 4times vs.\ GQA-8 at dense scales and 2times vs.\ GQA-4 at the MoE scale, and lowering training and prefill FLOPs by 7.9\% (1.7B, S{=}4096). On an industrial 30B-A3B MoE model trained on 500B tokens at 8K context, VHA outperforms both MLA and GQA-4 by +0.70\,/\,+2.60\,\% on a 12-benchmark base-model evaluation suite; in a separate matched 10B-A1B MoE comparison after long-context mid-training, VHA attains an average RULER score of 84.4 over 8K--128K versus 73.0 for GQA, while using two rather than four KV heads in the full-attention layers.

Lung nodules can be an alarming precursor to potential lung cancer. Missed nodule detections during chest radiograph analysis remains a common challenge among thoracic radiologists. In this work, we present a multi-task lung nodule detection algorithm for chest radiograph analysis. Unlike past approaches, our algorithm predicts a global-level label indicating nodule presence along with local-level labels predicting nodule locations using a Dual Head Network (DHN). We demonstrate the favorable nodule detection performance that our multi-task formulation yields in comparison to conventional methods. In addition, we introduce a novel Dual Head Augmentation (DHA) strategy tailored for DHN, and we demonstrate its significance in further enhancing global and local nodule predictions.

The Attention module finds common usage in language modeling, presenting distinct challenges within the broader scope of Natural Language Processing. Multi-Head Attention (MHA) employs an absolute positional encoding, which imposes limitations on token length and entails substantial memory consumption during the processing of embedded inputs. The current remedy proposed by researchers involves the utilization of relative positional encoding, similar to the approach adopted in Transformer-XL or Relative Multi-Head Attention (RMHA), albeit the employed architecture consumes considerable memory resources. To address these challenges, this study endeavors to refine MHA, leveraging relative positional encoding in conjunction with the Depth-Wise Convolutional Layer architecture, which promises heightened accuracy coupled with minimized memory usage. The proposed RCMHA framework entails the modification of two integral components: firstly, the application of the Depth-Wise Convolutional Layer to the input embedding, encompassing Query, Key, and Value parameters; secondly, the incorporation of Relative Positional Encoding into the attention scoring phase, harmoniously integrated with Scaled Dot-Product Attention. Empirical experiments underscore the advantages of RCMHA, wherein it exhibits superior accuracy, boasting a score of 0.572 in comparison to alternative attention modules such as MHA, Multi-DConv-Head Attention (MDHA), and RMHA. Concerning memory utilization, RMHA emerges as the most frugal, demonstrating an average consumption of 2.98 GB, surpassing RMHA which necessitates 3.5 GB.

Blind individuals face persistent challenges in last-mile navigation, including locating entrances, identifying obstacles, and navigating complex or cluttered spaces. Although wearable cameras are increasingly used in assistive systems, there has been no systematic, vantage-focused comparison to guide their design. This paper addresses that gap through a two-part investigation. First, we surveyed ten experienced blind cane users, uncovering navigation strategies, pain points, and technology preferences. Participants stressed the importance of multi-sensory integration, destination-focused travel, and assistive tools that complement (rather than replace) the cane's tactile utility. Second, we conducted controlled data collection with a blind participant navigating five real-world environments using synchronized head- and cane-mounted cameras, isolating vantage placement as the primary variable. To assess how each vantage supports spatial perception, we evaluated SLAM performance (for localization and mapping) and NeRF-based 3D reconstruction (for downstream scene understanding). Head-mounted sensors delivered superior localization accuracy, while cane-mounted views offered broader ground-level coverage and richer environmental reconstructions. A combined (head+cane) configuration consistently outperformed both. These results highlight the complementary strengths of different sensor placements and offer actionable guidance for developing hybrid navigation aids that are perceptive, robust, and user-aligned.

The introduction of ISO 12913-2:2018 has provided a framework for standardized data collection and reporting procedures for soundscape practitioners. A strong emphasis was placed on the use of calibrated head and torso simulators (HATS) for binaural audio capture to obtain an accurate subjective impression and acoustic measure of the soundscape under evaluation. To auralise the binaural recordings as recorded or at set levels, the audio stimuli and the headphone setup are usually calibrated with a HATS. However, calibrated HATS are too financially prohibitive for most research teams, inevitably diminishing the availability of the soundscape standard. With the increasing availability of soundscape binaural recording datasets, and the importance of cross-cultural validation of the soundscape ISO standards, e.g.\ via the Soundscape Attributes Translation Project (SATP), it is imperative to assess the suitability of cost-effective headphone calibration methods to maximise availability without severely compromising on accuracy. Hence, this study objectively examines an open-circuit voltage (OCV) calibration method in comparison to a calibrated HATS on various soundcard and headphone combinations. Preliminary experiments found that calibration with the OCV method differed significantly from the reference binaural recordings in sound pressure levels, whereas negligible differences in levels were observed with the HATS calibration.

Gaze estimation is a fundamental task in many applications of computer vision, human computer interaction and robotics. Many state-of-the-art methods are trained and tested on custom datasets, making comparison across methods challenging. Furthermore, existing gaze estimation datasets have limited head pose and gaze variations, and the evaluations are conducted using different protocols and metrics. In this paper, we propose a new gaze estimation dataset called ETH-XGaze, consisting of over one million high-resolution images of varying gaze under extreme head poses. We collect this dataset from 110 participants with a custom hardware setup including 18 digital SLR cameras and adjustable illumination conditions, and a calibrated system to record ground truth gaze targets. We show that our dataset can significantly improve the robustness of gaze estimation methods across different head poses and gaze angles. Additionally, we define a standardized experimental protocol and evaluation metric on ETH-XGaze, to better unify gaze estimation research going forward. The dataset and benchmark website are available at https://ait.ethz.ch/projects/2020/ETH-XGaze

Modern T2V/I2V generators synthesize people increasingly hard to distinguish from authentic footage, while current evaluation suites lag: legacy benchmarks target manipulation-based forgeries, and recent synthetic-video benchmarks prioritize scale over realistic human depiction. We introduce SynthForensics, a people-centric benchmark of 20{,}445 videos from 8 T2V and 7 I2V open-source generators, paired-source from FF++/DFD reals, two-stage human-validated, in four compression versions with full metadata. In our paired-comparison human study, raters prefer SynthForensics in 71--77% of head-to-head comparisons against each of nine existing synthetic-video benchmarks, while facial-quality metrics fall within the FF++/DFD baseline range. Across 15 detectors and three protocols, face-based methods drop 13--55 AUC points (mean 27) from FF++ to SynthForensics and a further 23 under aggressive compression; fine-tuning closes the gap at a backward cost on legacy benchmarks; training from scratch shows synthetic and manipulation features largely disjoint for most detectors. We release dataset, pipeline, and code.

Vision-Language-Action (VLA) policies commonly run Vision-Language Model (VLM) backbones with billions of parameters at every policy inference, which costs latency and energy. We revisit a decoupled alternative for multi-task manipulation: separate vision and language encoders whose representations condition a compact action head. We run a standardized comparison that varies the vision encoder, the language encoder, and the action head while holding the demonstrations, the training-step budget, the tasks, the evaluation protocol, and the measurement platform fixed, against seven VLA baselines. The resulting Decoupled Embodiment Model (DEM) combines a fine-tuned DINOv3 vision encoder, a frozen NeoBERT language encoder, and a MeanFlow head that generates an action chunk in one forward pass. On 18 RoboCasa tasks evaluated with held-out instruction paraphrases and randomized scenes, DEM reaches 55.6\% mean success against 56.9\% for GR00T N1.7 and 54.6\% for π_{0.5}, and on three real-robot tasks it reaches 66.0\% against 68.0\% for GR00T N1.7. On the same workstation, DEM needs 6.1\,ms per policy forward pass, a maximum throughput of 162.7 policy calls per second, and draws an estimated 2.07\,J of GPU energy per call, eight to seventeen times the throughput and six to fifteen times less energy than these VLM-backbone policies. Within this trained-task regime, DEM sits on the observed success--latency--energy frontier and provides a strong, efficient baseline for language-conditioned robot skills.

A prevailing view in robot learning is that simulation alone is not enough; effective sim-to-real transfer is widely believed to require at least some real-world data collection or task-specific fine-tuning to bridge the gap between simulated and physical environments. We challenge that assumption. With sufficiently large-scale and diverse simulated synthetic training data, we show that zero-shot transfer to the real world is not only possible, but effective for both static and mobile manipulation. We introduce MolmoBot-Engine, a fully open-source pipeline for procedural data generation across robots, tasks, and diverse simulated environments in MolmoSpaces. With it, we release MolmoBot-Data, a dataset of 1.8 million expert trajectories for articulated object manipulation and pick-and-place tasks. We train three policy classes: MolmoBot, a Molmo2-based multi-frame vision-language model with a flow-matching action head; MolmoBot-Pi0, which replicates the π_0 architecture to enable direct comparison; and MolmoBot-SPOC, a lightweight policy suitable for edge deployment and amenable to RL fine-tuning. We evaluate on two robotic platforms: the Franka FR3 for tabletop manipulation tasks and the Rainbow Robotics RB-Y1 mobile manipulator for door opening, drawer manipulation, cabinet interaction, and mobile pick-and-place. Without any real-world fine-tuning, our policies achieve zero-shot transfer to unseen objects and environments. On tabletop pick-and-place, MolmoBot achieves a success rate of 79.2% in real world evaluations across 4 settings, outperforming π_{0.5} at 39.2%. Our results demonstrate that procedural environment generation combined with diverse articulated assets can produce robust manipulation policies that generalize broadly to the real world. Technical Blog: https://allenai.org/blog/molmobot-robot-manipulation

Representation Autoencoders (RAEs) have shown distinct advantages in diffusion modeling on ImageNet by training in high-dimensional semantic latent spaces. In this work, we investigate whether this framework can scale to large-scale, freeform text-to-image (T2I) generation. We first scale RAE decoders on the frozen representation encoder (SigLIP-2) beyond ImageNet by training on web, synthetic, and text-rendering data, finding that while scale improves general fidelity, targeted data composition is essential for specific domains like text. We then rigorously stress-test the RAE design choices originally proposed for ImageNet. Our analysis reveals that scaling simplifies the framework: while dimension-dependent noise scheduling remains critical, architectural complexities such as wide diffusion heads and noise-augmented decoding offer negligible benefits at scale Building on this simplified framework, we conduct a controlled comparison of RAE against the state-of-the-art FLUX VAE across diffusion transformer scales from 0.5B to 9.8B parameters. RAEs consistently outperform VAEs during pretraining across all model scales. Further, during finetuning on high-quality datasets, VAE-based models catastrophically overfit after 64 epochs, while RAE models remain stable through 256 epochs and achieve consistently better performance. Across all experiments, RAE-based diffusion models demonstrate faster convergence and better generation quality, establishing RAEs as a simpler and stronger foundation than VAEs for large-scale T2I generation. Additionally, because both visual understanding and generation can operate in a shared representation space, the multimodal model can directly reason over generated latents, opening new possibilities for unified models.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

Instruction-following information retrieval (IF-IR) studies retrieval systems that must not only find documents relevant to a query, but also obey explicit user constraints such as required attributes, exclusions, or output preferences. However, most retrievers are trained primarily for semantic relevance and often fail to distinguish documents that match the topic from those that satisfy the instruction. We propose a dual-view data synthesis strategy based on polarity reversal: given a query, a document that is relevant under the instruction, and a hard negative that matches the query but violates the instruction, we prompt an LLM to generate a complementary instruction under which the two documents swap relevance labels. By presenting the same document pair under complementary instructions that invert their relevance labels, the training signal forces the retriever to reconsider the same candidate set through the instruction, rather than relying on fixed topical cues. On a 305M-parameter encoder, our method improves performance on the FollowIR benchmark by 45%, surpassing general-purpose embedding models of comparable or larger scale. Through head-to-head comparisons at matched data budgets, we further show that data diversity and instruction supervision play complementary roles: the former preserves general retrieval quality, while the latter improves instruction sensitivity. These results highlight the value of targeted data synthesis for building retrieval systems that are both broadly capable and instruction-aware.

A network of oscillators that synchronizes perfectly computes nothing further, so an attention architecture built from synchronization must locate its computation in structured departures from agreement. We introduce the Frustrated Synchronization Network (FSN), whose token states are phases on a torus and whose entire value pathway is one learned complex coupling kernel over harmonics and a one-step delay. Each component of the kernel is a frustration in the sense of the synchronization literature. The complex phases are static Kuramoto-Sakaguchi frustration angles, the signed harmonics are repulsive Daido components, and the delay term, which couples each token to the successors of the tokens it attends to, is algebraically identical to Kuramoto-Sakaguchi coupling whose frustration angle is the data's own transition, so next-token prediction is implemented as synchronization frustrated by the data. At matched one-million-parameter and training budgets on character-level text and code, the FSN's validation loss is below a tuned RoPE-SwiGLU transformer's at every epoch measured, and the comparison survives training the baseline to convergence: every thirty-epoch enwik8 seed finishes below the transformer's converged fifty-epoch loss of 1.611, and the FSN's completed fifty-epoch runs converge to 1.5953 +/- 0.0014. A variant with every feed-forward block replaced by mean-field coupling to learned collective modes, leaving no multilayer perceptron in the stack, tracks the transformer. On natural text the unfrustrated base layer falls behind the converged transformer at every copy depth, worst on long-range copy events; the kernel reverses the deficit at every depth of four and beyond. Headline comparisons are at the one-million-parameter scale; a scale ladder is complete through four million parameters with the advantage persisting, and remaining arms are marked as in progress.

Self-evolving agents are usually built around prescribed optimization pipelines: the framework decides how to gather evidence, revise a persistent artifact, select candidates, and stop. We ask whether this task-specific procedure remains necessary when a frontier model acts as the optimizer. We introduce Open-Ended Optimization (OEO), which keeps the objective, permitted interactions, resource budget, data boundary, and evaluation fixed while allowing the optimizer to compose the improvement process online. We compare OEO with two complementary prescribed approaches: SkillOpt, a staged pipeline with bounded edits, and GEPA, a reflective evolutionary search. Across 14 head-to-head comparisons over 8 benchmark-target-model settings, GPT-5.5-driven OEO records 12 wins, 1 tie, and 1 narrow loss of 0.21 percentage points. It uses a median 34.3 percent of SkillOpt's configured target-interaction token budget. A one-shot, zero-interaction control shows that the gains are not explained by a single prior-driven rewrite. However, delegation has a capability boundary: SkillOpt outperforms OEO with a medium optimizer, and a weak optimizer cannot operate through the unchanged OEO interface. In the fully instrumented OEO-SkillOpt pair, trajectory analysis further shows that prescription changes how optimization proceeds more consistently than it changes final behavior. Together, these findings recast prescribed pipelines as capability-dependent scaffolding: essential constraints remain external, but a sufficiently capable optimizer can compose the route from measurable feedback to persistent improvement.

Physicians now pose millions of clinical questions to AI tools each week, yet these tools are evaluated largely on hypothetical or exam-style questions, not those actually asked in practice. We report a blinded evaluation built on 620 Real-world Point-Of-Care Queries (Real-POCQi) submitted to the OpenEvidence (OE) platform by physicians spanning 30 specialties, as well as 187 questions from HealthBench. 149 practicing physicians across 36 states made head-to-head comparisons between answers from three frontier general-purpose models (Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5) and a specialized clinical tool (OE), with graders matched to each question's specialty. When comparing answers along five dimensions relevant to clinical decision support -- accuracy, clinical utility, source quality, verifiability, & completeness -- physicians scored the specialized tool highest on all axes; in the primary analysis on Real-POCQi, win differences (margins between win and loss rates) ranged from 25 to 39 percentage points (p<0.001). Results remained consistent in sensitivity analyses stratifying by citation display, answer length, OE-user status, and Real-POCQi versus HealthBench. In parallel, LLM judges were found to systematically differ from expert judges, though both generally agreed on the best model. These findings underscore two conclusions: (i) AI tool evaluations should reflect real-world query distributions and use expert judges that mirror the specialization defining modern medicine and (ii) the consistent advantage of the specialized tool over general-purpose models does not necessarily mean that the latter cannot serve similar purposes, but that targeted engineering and customization can yield meaningful gains in performance for its users. We release Real-POCQi as a public benchmark, as well as the prespecified statistical analysis for reproducing results of this study.

The Segment Anything Model 2 (SAM 2) has emerged as a powerful foundation model for object segmentation in both images and videos, paving the way for various downstream video applications. The crucial design of SAM 2 for video segmentation is its memory module, which prompts object-aware memories from previous frames for current frame prediction. However, its greedy-selection memory design suffers from the "error accumulation" problem, where an errored or missed mask will cascade and influence the segmentation of the subsequent frames, which limits the performance of SAM 2 toward complex long-term videos. To this end, we introduce SAM2Long, an improved training-free video object segmentation strategy, which considers the segmentation uncertainty within each frame and chooses the video-level optimal results from multiple segmentation pathways in a constrained tree search manner. In practice, we maintain a fixed number of segmentation pathways throughout the video. For each frame, multiple masks are proposed based on the existing pathways, creating various candidate branches. We then select the same fixed number of branches with higher cumulative scores as the new pathways for the next frame. After processing the final frame, the pathway with the highest cumulative score is chosen as the final segmentation result. Benefiting from its heuristic search design, SAM2Long is robust toward occlusions and object reappearances, and can effectively segment and track objects for complex long-term videos. Notably, SAM2Long achieves an average improvement of 3.0 points across all 24 head-to-head comparisons, with gains of up to 5.3 points in J&F on long-term video object segmentation benchmarks such as SA-V and LVOS. The code is released at https://github.com/Mark12Ding/SAM2Long.

Online reinforcement learning (RL) has been central to post-training language models, but its extension to diffusion models remains challenging due to intractable likelihoods. Recent works discretize the reverse sampling process to enable GRPO-style training, yet they inherit fundamental drawbacks, including solver restrictions, forward-reverse inconsistency, and complicated integration with classifier-free guidance (CFG). We introduce Diffusion Negative-aware FineTuning (DiffusionNFT), a new online RL paradigm that optimizes diffusion models directly on the forward process via flow matching. DiffusionNFT contrasts positive and negative generations to define an implicit policy improvement direction, naturally incorporating reinforcement signals into the supervised learning objective. This formulation enables training with arbitrary black-box solvers, eliminates the need for likelihood estimation, and requires only clean images rather than sampling trajectories for policy optimization. DiffusionNFT is up to 25times more efficient than FlowGRPO in head-to-head comparisons, while being CFG-free. For instance, DiffusionNFT improves the GenEval score from 0.24 to 0.98 within 1k steps, while FlowGRPO achieves 0.95 with over 5k steps and additional CFG employment. By leveraging multiple reward models, DiffusionNFT significantly boosts the performance of SD3.5-Medium in every benchmark tested.

Recent works propose complex multi-modal models that handle both time series and language, ultimately claiming high performance on complex tasks like time series reasoning and cross-modal question-answering. However, they skip evaluations of simple and important foundational tasks, which complex models should reliably master. They also lack direct, head-to-head comparisons with other popular approaches. So we ask a simple question: Can recent models even produce generic visual descriptions of time series data? In response, we propose three new tasks, posing that successful multi-modal models should be able to recognize, differentiate, and generate language descriptions of time series. We then create BEDTime, the first benchmark dataset to assess models on each task, comprising four datasets reformatted for these tasks across multiple modalities. Using BEDTime, we evaluate 13 state-of-the-art models, and find that (1) surprisingly, dedicated time series foundation models severely underperform, despite being designed for similar tasks, (2) vision-language models are quite capable, (3) language-only methods perform worst, despite many lauding their potential, and (4) all approaches are clearly fragile to a range of realistic robustness tests, indicating avenues for future work.

Reference listening is a common strategy in music production, but current comparison tools often obscure a key human judgment: deciding what should be compared. We present Diptych, an AI-assisted system that lets users define comparison scope across whole tracks or independently selected segments, while inspecting structured audio features and scope-specific AI interpretations. We evaluated Diptych in a within-participants study with 12 musicians, complemented by source-blinded ratings from four expert listeners. Participants used the system to surface additional differences, nine of ten of which received at least partial expert support, and reported good usability and greater clarity about possible next steps. These findings suggest that AI support for creative comparison should prioritize user-defined scope, inspectable evidence, and actionable guidance, while avoiding authoritative judgments that exceed what the evidence can support.

Data synthesis has been advocated as an important approach for utilizing data while protecting data privacy. A large number of tabular data synthesis algorithms (which we call synthesizers) have been proposed. Some synthesizers satisfy Differential Privacy, while others aim to provide privacy in a heuristic fashion. A comprehensive understanding of the strengths and weaknesses of these synthesizers remains elusive due to drawbacks in evaluation metrics and missing head-to-head comparisons of newly developed synthesizers that take advantage of diffusion models and large language models with state-of-the-art marginal-based synthesizers. In this paper, we present a systematic evaluation framework for assessing tabular data synthesis algorithms. Specifically, we examine and critique existing evaluation metrics, and introduce a set of new metrics in terms of fidelity, privacy, and utility to address their limitations. Based on the proposed metrics, we also devise a unified objective for tuning, which can consistently improve the quality of synthetic data for all methods. We conducted extensive evaluations of 8 different types of synthesizers on 12 real-world datasets and identified some interesting findings, which offer new directions for privacy-preserving data synthesis.

To advance Chinese financial natural language processing (NLP), we introduce BBT-FinT5, a new Chinese financial pre-training language model based on the T5 model. To support this effort, we have built BBT-FinCorpus, a large-scale financial corpus with approximately 300GB of raw text from four different sources. In general domain NLP, comprehensive benchmarks like GLUE and SuperGLUE have driven significant advancements in language model pre-training by enabling head-to-head comparisons among models. Drawing inspiration from these benchmarks, we propose BBT-CFLEB, a Chinese Financial Language understanding and generation Evaluation Benchmark, which includes six datasets covering both understanding and generation tasks. Our aim is to facilitate research in the development of NLP within the Chinese financial domain. Our model, corpus and benchmark are released at https://github.com/ssymmetry/BBT-FinCUGE-Applications. Our work belongs to the Big Bang Transformer (BBT), a large-scale pre-trained language model project.

Building high-quality large language models (LLMs) for enterprise Arabic applications remains challenging due to the limited availability of digitized Arabic data. In this work, we present a data synthesis and refinement strategy to help address this problem, namely, by leveraging synthetic data generation and human-in-the-loop annotation to expand our Arabic training corpus. We further present our iterative post training recipe that is essential to achieving state-of-the-art performance in aligning the model with human preferences, a critical aspect to enterprise use cases. The culmination of this effort is the release of a small, 7B, open-weight model that outperforms similarly sized peers in head-to-head comparisons and on Arabic-focused benchmarks covering cultural knowledge, instruction following, RAG, and contextual faithfulness.

LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about pop-up banner closure. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.

While generative video models have achieved remarkable fidelity and consistency, applying these capabilities to video editing remains a complex challenge. Recent research has explored motion controllability as a means to enhance text-to-video generation or image animation; however, we identify precise motion control as a promising yet under-explored paradigm for editing existing videos. In this work, we propose modifying video motion by directly editing sparse trajectories extracted from the input. We term the deviation between input and output trajectories a "motion edit" and demonstrate that this representation, when coupled with a generative backbone, enables powerful video editing capabilities. To achieve this, we introduce a pipeline for generating "motion counterfactuals", video pairs that share identical content but distinct motion, and we fine-tune a motion-conditioned video diffusion architecture on this dataset. Our approach allows for edits that start at any timestamp and propagate naturally. In a four-way head-to-head user study, our model achieves over 65 percent preference against prior work. Please see our project page: https://ryanndagreat.github.io/MotionV2V

We introduce CompareBench, a benchmark for evaluating visual comparison reasoning in vision-language models (VLMs), a fundamental yet understudied skill. CompareBench consists of 1000 QA pairs across four tasks: quantity (600), temporal (100), geometric (200), and spatial (100). It is derived from two auxiliary datasets that we constructed: TallyBench (2000 counting images with QA) and HistCaps (515 historical images with bilingual captions). We evaluate both closed-source APIs (OpenAI, Gemini, Claude) and open-source models (Qwen2.5-VL and Qwen3-VL series). Results show clear scaling trends but also reveal critical limitations: even the strongest models consistently fail at temporal ordering and spatial relations, and they often make mistakes in basic counting and geometric comparisons that are trivial for humans. These findings demonstrate that visual comparison remains a systematic blind spot for current VLMs. By providing controlled, diverse, and diagnostic evaluation, CompareBench establishes a foundation for advancing more reliable multimodal reasoning.

In contrastive learning, two views of an original image, generated by different augmentations, are considered a positive pair, and their similarity is required to be high. Similarly, two views of distinct images form a negative pair, with encouraged low similarity. Typically, a single similarity measure, provided by a lone projection head, evaluates positive and negative sample pairs. However, due to diverse augmentation strategies and varying intra-sample similarity, views from the same image may not always be similar. Additionally, owing to inter-sample similarity, views from different images may be more akin than those from the same image. Consequently, enforcing high similarity for positive pairs and low similarity for negative pairs may be unattainable, and in some cases, such enforcement could detrimentally impact performance. To address this challenge, we propose using multiple projection heads, each producing a distinct set of features. Our pre-training loss function emerges from a solution to the maximum likelihood estimation over head-wise posterior distributions of positive samples given observations. This loss incorporates the similarity measure over positive and negative pairs, each re-weighted by an individual adaptive temperature, regulated to prevent ill solutions. Our approach, Adaptive Multi-Head Contrastive Learning (AMCL), can be applied to and experimentally enhances several popular contrastive learning methods such as SimCLR, MoCo, and Barlow Twins. The improvement remains consistent across various backbones and linear probing epochs, and becomes more significant when employing multiple augmentation methods.

Comparative settings (e.g. pairwise choice, listwise ranking) have been adopted by a wide range of subjective studies for image quality assessment (IQA), as it inherently standardizes the evaluation criteria across different observers and offer more clear-cut responses. In this work, we extend the edge of emerging large multi-modality models (LMMs) to further advance visual quality comparison into open-ended settings, that 1) can respond to open-range questions on quality comparison; 2) can provide detailed reasonings beyond direct answers. To this end, we propose the Co-Instruct. To train this first-of-its-kind open-source open-ended visual quality comparer, we collect the Co-Instruct-562K dataset, from two sources: (a) LMM-merged single image quality description, (b) GPT-4V "teacher" responses on unlabeled data. Furthermore, to better evaluate this setting, we propose the MICBench, the first benchmark on multi-image comparison for LMMs. We demonstrate that Co-Instruct not only achieves 30% higher superior accuracy than state-of-the-art open-source LMMs, but also outperforms GPT-4V (its teacher), on both existing related benchmarks and the proposed MICBench. Our model is published at https://huggingface.co/q-future/co-instruct.

Foundation-model features are increasingly used to ask what information neural activity represents, often by comparing prediction gains between nested encoding models. We show that such multimodal contrasts can change sign when only the conditioning predictor is reconstructed. Using fMRI from the Natural Scenes Dataset, DINOv2 visual features, and MPNet embeddings of MS COCO captions and Localized Narratives, a caption-narrative contrast in the additional predictive contribution of vision favors narratives when one short caption is compared with a long narrative (+0.012/+0.015 in Places), but favors captions after approximate word-count matching (-0.031/-0.023). The shift occurs across every measured ROI in both subjects and is driven primarily by differences in language-only prediction. Comparable contrasts also survive removal of image-specific content-word identity in several ROIs. These results show that nested neural contrasts do not identify represented content by themselves: predictor construction is part of the experimental design, and matched controls are required for representational claims.

A manga reader meets a character on one page and knows them on sight a hundred pages later, without ever being handed a cast list. Re-identifying comic characters demands the same, open-set and sequential: pages arrive as a stream in reading order, new faces appear before anyone names them, and the cast is assembled as the story is read. Re:Cognize evaluates recognition as the story is read, not against a cast handed over in advance: four protocols on one query stream, from closed-set retrieval to a cast the model must build and grow itself. The surprise is where models fail. Recognising is close to solved: one reference image per character already ranks as well as a gallery built in advance. Knowing what to believe is not: a model that adds its own matches makes its cast worse, while the same growth with correct labels would gain over twenty points of top-1 accuracy. The bottleneck is acceptance, not vision, and one comparison decides it: an addition pays exactly when it is right more often than the cast already was on the queries it takes over. The comparison has nothing to fit, and measured on half of a new corpus it calls the other half correctly. ReCast puts it to work with nothing fitted on data: a cast sheet of one running average per character, grown only where the page itself vouches for a crop. It recovers a third to two thirds of what perfect labels would, depending on whether the cast starts from random examples or from first appearances. Re:Cognize measures whether a model can read along; ReCast is a cast that does. Our claims are on identity maintenance, recognising characters already met; the emergence of new ones is measured as a diagnostic under a fixed reference rule, and we propose no method for it.

In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them. Identifying which attention heads perform this synthesis matters for interpreting long-context model behavior. Yet existing detectors miss these heads by construction: they reward heads whose attended token matches the generated token, a literal-copy criterion that captures where a head reads but not what it writes through its output-value (OV) circuit, the very mechanism that carries non-literal retrieval. We introduce Logit-Contribution Scoring (LOCOS), a write-aware detector that scores each head by the projection of its OV-circuit output onto the answer-token unembedding direction, contrasting needle and off-needle source positions in a single forward pass. Across three model families (Qwen3, Gemma-3, OLMo-3.1), mean-ablating the top LOCOS heads on the NoLiMa non-literal retrieval benchmark collapses ROUGE-L at lower head counts than prior attention-based detections; on Qwen3-8B, ablating 50 heads drives ROUGE-L from 0.401 to 0.000 while the strongest baseline still retains 0.292. The selected heads are retrieval-specific: parametric recall and arithmetic reasoning stay at baseline under the same ablation. On Qwen3-8B, the same ablation also drops MuSiQue from 0.55 to 0.08 and BABI-Long from 0.62 to 0.20, while a random-heads control stays within 0.05 of baseline.

There is a multitude of novel generative models for open-domain conversational systems; however, there is no systematic evaluation of different systems. Systematic comparisons require consistency in experimental design, evaluation sets, conversational systems and their outputs, and statistical analysis. We lay out a protocol for the evaluation of conversational models using head-to-head pairwise comparison. We analyze ten recent models that claim state-of-the-art performance using a paired head-to-head performance (win-loss-tie) on five evaluation datasets. Our findings show that DialoGPT and Blender are superior systems using Bradley-Terry model and TrueSkill ranking methods. These findings demonstrate the feasibility of our protocol to evaluate conversational agents and evaluation sets. Finally, we make all code and evaluations publicly available for researchers to compare their model to other state-of-the-art dialog models.

While recent advancements in large multimodal models (LMMs) have significantly improved their abilities in image quality assessment (IQA) relying on absolute quality rating, how to transfer reliable relative quality comparison outputs to continuous perceptual quality scores remains largely unexplored. To address this gap, we introduce Compare2Score-an all-around LMM-based no-reference IQA (NR-IQA) model, which is capable of producing qualitatively comparative responses and effectively translating these discrete comparative levels into a continuous quality score. Specifically, during training, we present to generate scaled-up comparative instructions by comparing images from the same IQA dataset, allowing for more flexible integration of diverse IQA datasets. Utilizing the established large-scale training corpus, we develop a human-like visual quality comparator. During inference, moving beyond binary choices, we propose a soft comparison method that calculates the likelihood of the test image being preferred over multiple predefined anchor images. The quality score is further optimized by maximum a posteriori estimation with the resulting probability matrix. Extensive experiments on nine IQA datasets validate that the Compare2Score effectively bridges text-defined comparative levels during training with converted single image quality score for inference, surpassing state-of-the-art IQA models across diverse scenarios. Moreover, we verify that the probability-matrix-based inference conversion not only improves the rating accuracy of Compare2Score but also zero-shot general-purpose LMMs, suggesting its intrinsic effectiveness.

Radiological similarity depends on disease relationships and on fine details such as laterality, lobe, severity, size, and certainty. Broad biomedical similarity can overlook these qualifiers, particularly when several attributes vary together. We introduce QureRadEmbed, a 4B radiology-aware encoder trained with two complementary signals: RadSim supplies deterministic, attribute-decomposed ranking targets, while RadThought aligns reports with hierarchical evidence and reasoning descriptions. A three-stage curriculum combines these signals with report triplets, finding perturbations, and single- and cross-attribute contrasts. The final model achieves 0.996 mean ordering accuracy across ten controlled synthetic attributes and raises Spearman correlation with the designed joint-attribute targets from 0.501 to 0.976. On external findings-to-impression retrieval, Recall@1 reaches 10.5% on Open-I, 12.4% on testing XR, and 42.9% on testing CT, compared with 6.6%, 5.7%, and 31.4% for its backbone. Frozen embeddings support finding extraction with only 100 labeled testing-XR reports (macro-F1 0.481 versus 0.412 for the backbone). Whole-report comparison costs 8.8 seconds per 1,000 pairs in our benchmark, versus 2,755.1 seconds for the generative evaluator GREEN. Sentence-level comparison improves sensitivity to local discrepancies, although generative evaluation remains stronger on several expert-rated and subtle-error tasks. The results support reusable radiology-aware representations for search, structured report indexing, and efficient report comparison.

Opinion summarization focuses on generating summaries that reflect popular subjective information expressed in multiple online reviews. While generated summaries offer general and concise information about a particular hotel or product, the information may be insufficient to help the user compare multiple different choices. Thus, the user may still struggle with the question "Which one should I pick?" In this paper, we propose the comparative opinion summarization task, which aims at generating two contrastive summaries and one common summary from two different candidate sets of reviews. We develop a comparative summarization framework CoCoSum, which consists of two base summarization models that jointly generate contrastive and common summaries. Experimental results on a newly created benchmark CoCoTrip show that CoCoSum can produce higher-quality contrastive and common summaries than state-of-the-art opinion summarization models. The dataset and code are available at https://github.com/megagonlabs/cocosum

With the advent of highly capable instruction-tuned neural language models, benchmarking in natural language processing (NLP) is increasingly shifting towards pairwise comparison leaderboards, such as LMSYS Arena, from traditional global pointwise scores (e.g., GLUE, BIG-bench, SWE-bench). This paper empirically investigates the strengths and weaknesses of both global scores and pairwise comparisons to aid decision-making in selecting appropriate model evaluation strategies. Through computational experiments on synthetic and real-world datasets using standard global metrics and the popular Bradley-Terry model for pairwise comparisons, we found that while global scores provide more reliable overall rankings, they can underestimate strong models with rare, significant errors or low confidence. Conversely, pairwise comparisons are particularly effective for identifying strong contenders among models with lower global scores, especially where quality metrics are hard to define (e.g., text generation), though they require more comparisons to converge if ties are frequent. Our code and data are available at https://github.com/HSPyroblast/srw-ranking under a permissive license.

Automated summary quality assessment falls into two categories: reference-based and reference-free. Reference-based metrics, historically deemed more accurate due to the additional information provided by human-written references, are limited by their reliance on human input. In this paper, we hypothesize that the comparison methodologies used by some reference-based metrics to evaluate a system summary against its corresponding reference can be effectively adapted to assess it against its source document, thereby transforming these metrics into reference-free ones. Experimental results support this hypothesis. After being repurposed reference-freely, the zero-shot BERTScore using the pretrained DeBERTa-large-MNLI model of <0.5B parameters consistently outperforms its original reference-based version across various aspects on the SummEval and Newsroom datasets. It also excels in comparison to most existing reference-free metrics and closely competes with zero-shot summary evaluators based on GPT-3.5.

Many vision applications ask several independent, forced-choice questions about the same image. Visual Jev encodes the image and public context once, executes isolated question suffixes as a batch, and reads candidate probabilities from the backbone's language-model head. Across four benchmarks, answer-supervised post-training raises equal-weight macro accuracy from 70.6% to 76.1%, with the gain concentrated on the two task families represented in training. At N=32 questions per image, shared batched execution is 8.9x faster in warm amortized time than independent serial execution and remains 3.4x faster than an already-batched baseline that recomputes the prefix, at the cost of higher peak memory. A matched typed-head control offers no consistent accuracy advantage over the language-model-head readout. The supported design is therefore simple: adapt the backbone for quality, retain the existing readout, and share execution for efficiency.

We present HEAD-QA, a multi-choice question answering testbed to encourage research on complex reasoning. The questions come from exams to access a specialized position in the Spanish healthcare system, and are challenging even for highly specialized humans. We then consider monolingual (Spanish) and cross-lingual (to English) experiments with information retrieval and neural techniques. We show that: (i) HEAD-QA challenges current methods, and (ii) the results lag well behind human performance, demonstrating its usefulness as a benchmark for future work.

Locating relevant evidence in long inputs is a core capability of Multimodal Large Language Models. Prior studies attribute this capability to retrieval heads, a sparse set of attention heads that transfer information from the context to the output. However, these analyses are confined to a limited set of modalities and the corresponding models. Moreover, how retrieval heads can be leveraged to enhance the model calls for further exploration. In this paper, we ask whether this mechanism expands together with the input modalities of the model. Taking audio-visual input as a newly added modality, we construct paired inputs in which the same answer is either spoken or shown on screen. With the evidence content fixed, retrieval heads remain sparse in the new modality and are shared to a large extent across modalities. Across 18 language, vision-language, and omni-modal checkpoints, omni-modal models largely preserve the retrieval heads of their backbones. Ablating these heads eliminates nearly all audio-visual retrieval. The heads shared across modalities make up fewer than 2\% of all heads, yet ablating them also degrades document understanding, spatial perception, and visual mathematics. Building on these findings, we turn retrieval heads into a target of optimization by training their attention with a ranking objective. This raises the page Recall@1 of Qwen3-VL-8B on MMDocIR by 6.03 points and improves document retrieval across three model families. Functional attention heads can thus serve both as a lens on multimodal models and as a target for improving them.

A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term. We measure the second. The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate shipped as gold, and a reviewer gold from independent expert re-annotation of a sample. Using it, we hold the scored output fixed and exchange the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other (95% CIs [3.23, 6.27] and [0.26, 3.37]), and by 7.55 in the worst language. The comparison between two outputs moves as well: the paired interaction between reference and system is +2.95 points (CI [+1.97, +3.87]), survives correction for multiple testing, and changes one language's margin outright. The cause is an undocumented aggregation default that usually kept one annotator when adjudication did not fire, making the shipped reference a partial copy of an output being scored. We report the resulting reference-sensitivity band, show how to compute one from any retained annotation record, and argue that the quantity belongs beside the score.

Large vision-language models (LVLMs) can process interleaved text and multiple images in a single context, but how they internally identify the image relevant to a question remains poorly understood. We study this grounding step as in-context image retrieval and introduce ICIR-MCQ, a controlled probe for head-level analysis of image retrieval. Using each head's attention mass over image-token spans, we first identify sparse pointer heads that reliably point to the target image. These heads transfer beyond the probe: a single pointer head is competitive with external vision-language retrievers on a naturalistic multi-image retrieval benchmark, showing that LVLMs contain a strong internal image retrieval signal. Attention alone, however, shows where a head reads from, not what its output passes on to later layers, so we introduce a complementary output-side analysis. For each head, we train a lightweight image retriever on its value-weighted output and use it to identify writer heads, whose outputs contain enough information to identify the target image. Surprisingly, the pointer and writer criteria select substantially different head sets, with only 35 to 47 heads overlapping among the top-100 heads under each criterion. Targeted set-partition ablations show that the heads most important for standard multi-image inference are concentrated in this pointer-writer intersection, while heads selected by only one criterion show less consistent separation from random-head controls. The most consistent causal effects arise from heads selected by both the attention-based pointer and output-side writer criteria. Code and data: https://github.com/parkquasar/Image-Retrieval-Heads.

Reinforcement learning from verifiable rewards (e.g. GRPO) is the engine behind today's reasoning models, yet it grades only the final answer. On hard problems this trains models to write more rather than to think better, since the trace itself is never graded and no label for good thinking exists. We introduce Agon, which makes two competing models each other's graders. Both attempt the same problem; in alternating roles, one drafts a solution and the other reads it while solving, and each is rewarded for out-solving the other. To win, a model must out-reason a rival that has seen its work, so reasoning is judged implicitly during training, with no process labels and no reward model. Because both models are optimized, each faces a progressively stronger rival, which single-model RL cannot provide. The two need only be comparably strong and behaviorally different. At inference the pair deploys as it trains, a two-stage cascade in which one model drafts and the other answers after reading the draft. On the hard split of DeepMath with Qwen3, this doubles GRPO's pass@1, roughly eight times the gain of an untrained Mixture-of-Agents pass over the same base. The ordering replicates on competitive-programming code and across model families (Qwen3.5, Gemma 4). For now the models talk in text; the next step is to let them reason together in latent space.

Comparing two images in terms of Commonalities and Differences (CaD) is a fundamental human capability that forms the basis of advanced visual reasoning and interpretation. It is essential for the generation of detailed and contextually relevant descriptions, performing comparative analysis, novelty detection, and making informed decisions based on visual data. However, surprisingly, little attention has been given to these fundamental concepts in the best current mimic of human visual intelligence - Large Multimodal Models (LMMs). We develop and contribute a new two-phase approach CaD-VI for collecting synthetic visual instructions, together with an instruction-following dataset CaD-Inst containing 349K image pairs with CaD instructions collected using CaD-VI. Our approach significantly improves the CaD spotting capabilities in LMMs, advancing the SOTA on a diverse set of related tasks by up to 17.5%. It is also complementary to existing difference-only instruction datasets, allowing automatic targeted refinement of those resources increasing their effectiveness for CaD tuning by up to 10%. Additionally, we propose an evaluation benchmark with 7.5K open-ended QAs to assess the CaD understanding abilities of LMMs.