Data curation is the problem of how to collect and organize samples into a dataset that supports efficient learning. Despite the centrality of the task, little work has been devoted towards a large-scale, systematic comparison of various curation methods. In this work, we take steps towards a formal evaluation of data curation strategies and introduce SELECT, the first large-scale benchmark of curation strategies for image classification. In order to generate baseline methods for the SELECT benchmark, we create a new dataset, ImageNet++, which constitutes the largest superset of ImageNet-1K to date. Our dataset extends ImageNet with 5 new training-data shifts, each approximately the size of ImageNet-1K itself, and each assembled using a distinct curation strategy. We evaluate our data curation baselines in two ways: (i) using each training-data shift to train identical image classification models from scratch (ii) using the data itself to fit a pretrained self-supervised representation. Our findings show interesting trends, particularly pertaining to recent methods for data curation such as synthetic data generation and lookup based on CLIP embeddings. We show that although these strategies are highly competitive for certain tasks, the curation strategy used to assemble the original ImageNet-1K dataset remains the gold standard. We anticipate that our benchmark can illuminate the path for new methods to further reduce the gap. We release our checkpoints, code, documentation, and a link to our dataset at https://github.com/jimmyxu123/SELECT.
Scene text recognition (STR) suffers from the challenges of either less realistic synthetic training data or the difficulty of collecting sufficient high-quality real-world data, limiting the effectiveness of trained STR models. Meanwhile, despite producing holistically appealing text images, diffusion-based text image generation methods struggle to generate accurate and realistic instance-level text on a large scale. To tackle this, we introduce TextSSR: a novel framework for Synthesizing Scene Text Recognition data via a diffusion-based universal text region synthesis model. It ensures accuracy by focusing on generating text within a specified image region and leveraging rich glyph and position information to create the less complex text region compared to the entire image. Furthermore, we utilize neighboring text within the region as a prompt to capture real-world font styles and layout patterns, guiding the generated text to resemble actual scenes. Finally, due to its prompt-free nature and capability for character-level synthesis, TextSSR enjoys a wonderful scalability and we construct an anagram-based TextSSR-F dataset with 0.4 million text instances with complexity and realism. Experiments show that models trained on added TextSSR-F data exhibit better accuracy compared to models trained on 4 million existing synthetic data. Moreover, its accuracy margin to models trained fully on a real-world dataset is less than 3.7%, confirming TextSSR's effectiveness and its great potential in scene text image synthesis. Our code is available at https://github.com/YesianRohn/TextSSR.
In real-world recommendation scenarios, users engage with items through various types of behaviors. Leveraging diversified user behavior information for learning can enhance the recommendation of target behaviors (e.g., buy), as demonstrated by recent multi-behavior methods. The mainstream multi-behavior recommendation framework consists of two steps: fusion and prediction. Recent approaches utilize graph neural networks for multi-behavior fusion and employ multi-task learning paradigms for joint optimization in the prediction step, achieving significant success. However, these methods have limited perspectives on multi-behavior fusion, which leads to inaccurate capture of user behavior patterns in the fusion step. Moreover, when using multi-task learning for prediction, the relationship between the target task and auxiliary tasks is not sufficiently coordinated, resulting in negative information transfer. To address these problems, we propose a novel multi-behavior recommendation framework based on the combinatorial optimization perspective, named COPF. Specifically, we treat multi-behavior fusion as a combinatorial optimization problem, imposing different constraints at various stages of each behavior to restrict the solution space, thus significantly enhancing fusion efficiency (COGCN). In the prediction step, we improve both forward and backward propagation during the generation and aggregation of multiple experts to mitigate negative transfer caused by differences in both feature and label distributions (DFME). Comprehensive experiments on three real-world datasets indicate the superiority of COPF. Further analyses also validate the effectiveness of the COGCN and DFME modules. Our code is available at https://github.com/1918190/COPF.
Visual generative abductive learning studies jointly training symbol-grounded neural visual generator and inducing logic rules from data, such that after learning, the visual generation process is guided by the induced logic rules. A major challenge for this task is to reduce the time cost of logic abduction during learning, an essential step when the logic symbol set is large and the logic rule to induce is complicated. To address this challenge, we propose a pre-training method for obtaining meta-rule selection policy for the recently proposed visual generative learning approach AbdGen [Peng et al., 2023], aiming at significantly reducing the candidate meta-rule set and pruning the search space. The selection model is built based on the embedding representation of both symbol grounding of cases and meta-rules, which can be effectively integrated with both neural model and logic reasoning system. The pre-training process is done on pure symbol data, not involving symbol grounding learning of raw visual inputs, making the entire learning process low-cost. An additional interesting observation is that the selection policy can rectify symbol grounding errors unseen during pre-training, which is resulted from the memorization ability of attention mechanism and the relative stability of symbolic patterns. Experimental results show that our method is able to effectively address the meta-rule selection problem for visual abduction, boosting the efficiency of visual generative abductive learning. Code is available at https://github.com/future-item/metarule-select.
Video Large Language Models (Video LLMs) have achieved significant success by leveraging a two-stage paradigm: pretraining on large-scale video-text data for vision-language alignment, followed by supervised fine-tuning (SFT) for task-specific capabilities. However, existing approaches struggle with temporal reasoning due to weak temporal correspondence in the data and reliance on the next-token prediction paradigm during training. To address these limitations, we propose TEMPLE (TEMporal Preference Learning), a systematic framework that enhances Video LLMs' temporal reasoning capabilities through Direct Preference Optimization (DPO). To facilitate this, we introduce an automated preference data generation pipeline that systematically constructs preference pairs by selecting videos that are rich in temporal information, designing video-specific perturbation strategies, and finally evaluating model responses on clean and perturbed video inputs. Our temporal alignment features two key innovations: curriculum learning which that progressively increases perturbation difficulty to improve model robustness and adaptability; and "Pre-SFT Alignment'', applying preference optimization before instruction tuning to prioritize fine-grained temporal comprehension. Extensive experiments demonstrate that our approach consistently improves Video LLM performance across multiple benchmarks with a relatively small set of self-generated DPO data. We further analyze the transferability of DPO data across architectures and the role of difficulty scheduling in optimization. Our findings highlight our TEMPLE as a scalable and efficient complement to SFT-based methods, paving the way for developing reliable Video LLMs. Code is available at https://github.com/lscpku/TEMPLE.
The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss are prone to over-optimization, losing generalizability to unseen input distributions. In this paper, we study the cause of over-optimization in RM training and its downstream effects on the RLHF procedure, accentuating the importance of distributional robustness of RMs in unseen data. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Then, we propose batch-wise sum-to-zero regularization (BSR) to enforce zero-centered reward sum per batch, constraining the rewards with extreme magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness. Subsequently, we compare the plain BT model and BSR on RLHF training and empirically show that robust RMs better align the policy to the gold preference model. Finally, we apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale by adding more than 5% in complex preference prediction tasks. By conducting RLOO training with 8B RMs, AlpacaEval 2.0 reduces generation length by 40% while adding a 7% increase in win rate, further highlighting that robustness in RMs induces robustness in RLHF training. We release the code, data, and models: https://github.com/LinkedIn-XFACT/RM-Robustness.
Visual Place Recognition (VPR) is aimed at predicting the location of a query image by referencing a database of geotagged images. For VPR task, often fewer discriminative local regions in an image produce important effects while mundane background regions do not contribute or even cause perceptual aliasing because of easy overlap. However, existing methods lack precisely modeling and full exploitation of these discriminative regions. In this paper, we propose the Focus on Local (FoL) approach to stimulate the performance of image retrieval and re-ranking in VPR simultaneously by mining and exploiting reliable discriminative local regions in images and introducing pseudo-correlation supervision. First, we design two losses, Extraction-Aggregation Spatial Alignment Loss (SAL) and Foreground-Background Contrast Enhancement Loss (CEL), to explicitly model reliable discriminative local regions and use them to guide the generation of global representations and efficient re-ranking. Second, we introduce a weakly-supervised local feature training strategy based on pseudo-correspondences obtained from aggregating global features to alleviate the lack of local correspondences ground truth for the VPR task. Third, we suggest an efficient re-ranking pipeline that is efficiently and precisely based on discriminative region guidance. Finally, experimental results show that our FoL achieves the state-of-the-art on multiple VPR benchmarks in both image retrieval and re-ranking stages and also significantly outperforms existing two-stage VPR methods in terms of computational efficiency. Code and models are available at https://github.com/chenshunpeng/FoL
While large reasoning models demonstrate strong performance on complex tasks, they lack the ability to adjust reasoning token usage based on task difficulty. This often leads to the "overthinking" problem -- excessive and unnecessary reasoning -- which, although potentially mitigated by human intervention to control the token budget, still fundamentally contradicts the goal of achieving fully autonomous AI. In this work, we propose Adaptive Reasoning Model (ARM), a reasoning model capable of adaptively selecting appropriate reasoning formats based on the task at hand. These formats include three efficient ones -- Direct Answer, Short CoT, and Code -- as well as a more elaborate format, Long CoT. To train ARM, we introduce Ada-GRPO, an adaptation of Group Relative Policy Optimization (GRPO), which addresses the format collapse issue in traditional GRPO. Ada-GRPO enables ARM to achieve high token efficiency, reducing tokens by an average of 30%, and up to 70%, while maintaining performance comparable to the model that relies solely on Long CoT. Furthermore, not only does it improve inference efficiency through reduced token generation, but it also brings a 2x speedup in training. In addition to the default Adaptive Mode, ARM supports two additional reasoning modes: 1) Instruction-Guided Mode, which allows users to explicitly specify the reasoning format via special tokens -- ideal when the appropriate format is known for a batch of tasks. 2) Consensus-Guided Mode, which aggregates the outputs of the three efficient formats and resorts to Long CoT in case of disagreement, prioritizing performance with higher token usage.
This study presents a novel approach for decomposing urban water demand patterns using Skewed Gaussian Distributions (SGD) to derive behavioral insights and support operational planning. Hourly demand profiles contain critical information for both long-term infrastructure design and daily operations, influencing network pressures, water quality, energy consumption, and overall reliability. By breaking down each daily demand curve into a baseline component and distinct peak components, the proposed SGD method characterizes each peak with interpretable parameters, including peak amplitude, timing (mean), spread (duration), and skewness (asymmetry), thereby reconstructing the observed pattern and uncovering latent usage dynamics. This detailed peak-level decomposition enables both operational applications, e.g. anomaly and leakage detection, real-time demand management, and strategic analyses, e.g. identifying behavioral shifts, seasonal influences, or policy impacts on consumption patterns. Unlike traditional symmetric Gaussian or purely statistical time-series models, SGDs explicitly capture asymmetric peak shapes such as sharp morning surges followed by gradual declines, improving the fidelity of synthetic pattern generation and enhancing the detection of irregular consumption behavior. The method is demonstrated on several real-world datasets, showing that SGD outperforms symmetric Gaussian models in reconstruction accuracy, reducing root-mean-square error by over 50% on average, while maintaining physical interpretability. The SGD framework can also be used to construct synthetic demand scenarios by designing daily peak profiles with chosen characteristics. All implementation code is publicly available at: https://github.com/Relkayam/water-demand-decomposition-sgd
Reinforcement learning (RL) has shown great effectiveness for fine-tuning large language models (LLMs) using tasks that are challenging yet easily verifiable, such as math reasoning or code generation. However, extending this success to visual perception in vision-language models (VLMs) has been impeded by the scarcity of vision-centric tasks that are simultaneously challenging and unambiguously verifiable. To this end, we introduce ViCrit (Visual Caption Hallucination Critic), an RL proxy task that trains VLMs to localize a subtle, synthetic visual hallucination injected into paragraphs of human-written image captions. Starting from a 200-word captions, we inject a single, subtle visual description error-altering a few words on objects, attributes, counts, or spatial relations-and task the model to pinpoint the corrupted span given the image and the modified caption. This formulation preserves the full perceptual difficulty while providing a binary, exact-match reward that is easy to compute and unambiguous. Models trained with the ViCrit Task exhibit substantial gains across a variety of VL benchmarks. Crucially, the improvements transfer beyond natural-image training data to abstract image reasoning and visual math, showing promises of learning to perceive rather than barely memorizing seen objects. To facilitate evaluation, we further introduce ViCrit-Bench, a category-balanced diagnostic benchmark that systematically probes perception errors across diverse image domains and error types. Together, our results demonstrate that fine-grained hallucination criticism is an effective and generalizable objective for enhancing visual perception in VLMs.
Recently, continuous normalizing flows (CNFs) and diffusion models (DMs) have been studied using the unified theoretical framework. Although such models can generate high-quality data points from a noise distribution, the sampling demands multiple iterations to solve an ordinary differential equation (ODE) with high computational complexity. Most existing methods focus on reducing the number of time steps during the sampling process to improve efficiency. In this work, we explore a complementary direction in which the quality-complexity tradeoff can be dynamically controlled in terms of time steps and in the length of the neural network. We achieve this by rewiring the blocks in the transformer-based architecture to solve an inner discretized ODE w.r.t. its length. Then, we employ time- and length-wise consistency terms during flow matching training, and as a result, the sampling can be performed with an arbitrary number of time steps and transformer blocks. Unlike others, our $\textrm{ODE}_t \left(\textrm{ODE}_l \right)$ approach is solver-agnostic in time dimension and decreases both latency and memory usage. Compared to the previous state of the art, image generation experiments on CelebA-HQ and ImageNet show a latency reduction of up to $3\times$ in the most efficient sampling mode, and a FID score improvement of up to $3.5$ points for high-quality sampling. We release our code and model weights with fully reproducible experiments.
지연 가돌리늄 조영 증강(LGE) 심장 MRI는 심방세동 절제술 결과와 관련된 영상 지표인 좌심방(LA) 흉터를 비침습적으로 평가할 수 있게 한다. 그러나 낮은 대비, 얇은 심방벽, 부족한 레이블, 극심한 클래스 불균형으로 인해 자동 흉터 분할은 어렵다. 조건부 확산은 반복적인 마스크 생성을 제공하지만, 희소한 3D LA 흉터 분할에 대한 적합성은 아직 불분명하다. 이러한 공백을 해결하기 위해 동적 인코더로 조건화된 3D 사인파 UNet을 사용하여 깨끗한 마스크를 예측하는 조건부 마스크 확산(CMDiff)을 제안한다. 흉터 레이블이 있는 공개 LASCARQS 피험자 60명에 대해 폴드당 학습 48건과 홀드아웃 12건으로 5겹 교차 검증을 수행하여, 동일한 사례에서 CMDiff를 결정론적 기준 모델과 비교했다. CMDiff는 흉터 DSC 0.47 $\pm$ 0.08, HD95 12.89 $\pm$ 6.39 mm, ASD 1.91 $\pm$ 0.96 mm를 달성했으며, 결정론적 기준 모델은 각각 0.51 $\pm$ 0.08, 10.86 $\pm$ 5.39 mm, 1.58 $\pm$ 1.03 mm였다. 동일한 구현 하에서 깨끗한 마스크 예측은 본 연구에서 평가한 노이즈 예측 구성보다 상당히 우수한 성능을 보였다. 따라서 CMDiff는 기술적으로 실현 가능했으나 분할 성능을 향상시키지는 못했다. DSC는 더 낮았고 HD95와 ASD는 모두 더 높았다. 그러므로 직접 마스크 확산은 분석적 대안을 제공하지만, 이 과제에서 경쟁력 있는 성능 향상은 아니다. 소스 코드는 https://github.com/jizhang02/CMDiff 에 공개되어 있다.
스트리밍 영상 이해는 모델이 시간적으로 입력되는 프레임을 처리할 뿐 아니라, 증강현실(AR) 안경과 같은 현실적인 응용을 위해 사용자의 의도를 예측할 것을 요구한다. 기존 스트리밍 벤치마크는 시간적 추론을 평가하지만, 멀티모달 대규모 언어 모델(MLLM)이 스트리밍 환경에서 사람의 시선 신호를 해석하거나 활용할 수 있는지를 측정하는 벤치마크는 없다. 이러한 공백을 메우기 위해, 우리는 MLLM이 스트리밍 영상에서 시간적 및 능동적 추론을 위해 시선을 얼마나 효과적으로 활용하는지 평가하도록 설계된 최초의 벤치마크 StreamGaze를 소개한다. StreamGaze는 스트리밍 영상 이해를 포괄적으로 평가하는 시선 기반의 과거, 현재, 능동적 과제를 도입한다. 과거 및 현재 과제는 질의 시점에 이용 가능한 관측에 기반한 즉각적인 응답을 요구하는 반면, 능동적 과제는 모델이 입력 프레임을 지속적으로 처리하고 적절한 미래 시점에 응답할 것을 요구한다. StreamGaze를 구축하기 위해, 우리는 응시 추출, 영역별 시각적 프롬프팅, 스캔패스 구성을 통해 1인칭 시점 영상을 원시 시선 궤적과 정렬하는 시선-영상 질의응답(QA) 생성 파이프라인을 개발한다. 이 파이프라인은 사람의 지각 역학을 반영하는 시공간적으로 그라운딩된 QA 쌍을 생성한다. 모든 StreamGaze 과제에서 최신 MLLM과 사람의 성능 사이에 상당한 격차가 관찰되며, 이는 시선 기반 시간적 추론, 의도 모델링, 능동적 예측의 핵심적 한계를 부각한다. 또한 시선 프롬프팅 전략, 추론 행동, 과제별 실패 양상에 대한 상세한 분석을 제공하여 현재의 한계와 향후 연구 방향에 대한 통찰을 제시한다. 시선 기반 스트리밍 영상 이해 연구를 지원하기 위해 모든 데이터와 코드를 공개한다.
전 세계 태양광(PV) 발전 용량은 빠르게 증가하고 있으며, 이는 배출량과 화석 연료 의존도를 줄이는 데 기여한다. 그러나 태양광 발전소의 최적 입지를 선정하는 일은 지리적, 도시적, 인프라적 요구를 예측해야 하므로 여전히 어렵다. 현재의 방법은 직접 선별한 지구물리 레이어와 태양광 자원 제품을 사용하는 다기준 의사결정 분석(MCDA)에 의존하지만, 이러한 전략은 지리적으로 이식하기 어렵고 태양광 부지의 실질적 요구 사항을 반영하지 못한다. 본 연구에서는 위치의 지구물리적 속성을 나타내는 밀집 벡터인 지리공간 임베딩을 사용하여 기존 태양광 부지 38,591곳과 그 부지가 건설된 토지를 분석하고, 잠재적인 태양광 적합성 신호를 학습한다. 일련의 선형 및 비선형 모델을 평가하여 정확도와 대규모 배포 가능성을 고려해 랜덤 포레스트와 다층 퍼셉트론의 소프트 보팅 앙상블(RF+MLP 앙상블)을 선택한다. 13개국의 향후 설치 사례에 대해 검증한 결과, 이 앙상블은 기존 태양광 부지를 무작위 토지와 중앙값 ROC-AUC 0.90으로 구분한다. 발표된 MCDA 전략과 비교하면 모든 국가에서 ROC-AUC가 0.06~0.30 증가한다. 또한 이 순위는 실제 배치를 집중적으로 반영하여, 가장 적합한 상위 20%의 토지가 추정 발전량(연간 에너지 생산량)의 84%와 설치 용량(잠재 출력)의 67%를 차지한다. 표준 입지 제약을 적용한 후에도 전 세계 토지의 5.6%(농경지를 허용하면 9.8%)가 태양광에 매우 적합한 것으로 남는다. 전 지구 10 m 해상도 태양광 적합성 데이터셋과 코드는 https://github.com/Abdullahr29/Global-Solar-Siting 에서 확인할 수 있다.
UI 스크린샷을 통해 UX(User Experience)를 진단하는 UXBench 벤치마크를 제안하고, 강화학습(reinforcement learning) 기반의 새로운 보상 라우팅 메커니즘을 적용한 UI-UX 모델이 기존 최상위 모델보다 뛰어난 성능을 보임을 확인했습니다.
텍스트-이미지 디퓨전 모델에서 텍스트 설명을 사용하여 자연 이미지를 편집하는 것은, 특히 일관된 생성을 달성하고 복잡한 비강체 객체를 처리하는 데 있어 여전히 중요한 과제입니다. 기존 방법들은 텍스처와 정체성을 보존하는 데 어려움을 겪고, 많은 파인튜닝을 요하며, 배경 세부 정보를 유지하면서 특정 공간 영역이나 객체를 편집하는 데 있어 한계를 보입니다. 본 논문은 복잡하고 비강체인 실제 이미지 편집을 위한 새로운 제로샷 프레임워크인 문맥 보존 적응형 조작(Context-Preserving Adaptive Manipulation, CPAM)을 제안합니다. 구체적으로, 객체와 배경을 효과적으로 보존하고 독립적으로 제어하기 위해 셀프-어텐션 메커니즘을 조정하는 보존 적응 모듈을 제안합니다. 이는 마스크 가이던스 기법을 사용하여 편집 과정에서 배경이 왜곡되지 않도록 유지하면서 객체의 형태, 텍스처 및 정체성이 보존되도록 보장합니다. 또한, 크로스-어텐션 메커니즘의 컨디셔닝 과정에서 원하지 않는 수정 영역에 대한 간섭을 완화하기 위해 국소화 추출 모듈을 개발했습니다. 또한, 간단한 방식으로 다양한 이미지 조작 작업을 용이하게 하는 다양한 마스크 가이던스 전략을 도입했습니다. CPAM은 SD1.5, SD2.1, SDXL을 포함한 여러 디퓨전 백본과 매끄럽게 통합될 수 있으며, 서로 다른 모델 아키텍처에 걸쳐 강력한 일반화 성능을 입증합니다. 실제 이미지 편집을 위해 특별히 설계된 견고한 벤치마크 데이터셋인 새로 구축한 이미지 조작 벤치마크(Image Manipulation BenchmArk, IMBA)에 대한 광범위한 실험을 통해, 제안하는 방법이 기존 최첨단 편집 기술을 능가하며 인간 평가자들 사이에서 선호되는 선택임을 입증했습니다. 소스 코드와 데이터는 프로젝트 페이지에서 공개될 예정입니다: https://vdkhoi20.github.io/CPAM
Doctor agents are moving beyond single-turn answer generation toward evolving clinical decision systems. Within an outpatient episode, they acquire evidence, use examination and consultation resources, and decide when to finalize a diagnosis and management plan. Across episodes, their behavior may change through memory, retrieval, reflection, or other update mechanisms. Current evaluations only partially cover this setting. Fixed-input medical QA benchmarks score final answers from complete inputs, whereas many interactive benchmarks still focus on individual encounters or fixed runs, providing limited support for evaluating how episode-level decisions interact with cross-episode experience. We introduce MedEvoEval, an executable longitudinal evaluation framework based on action-gated simulated outpatient episodes. Each source case is converted into role-specific patient, examination, and manager views; evidence is revealed only through valid actions; and each episode records a structured trace that links observations, actions, final outputs, manager scores, and optional experience write-back. We release a runnable E&D artifact with 700 processed episodes, provenance notes, schemas, an episode runner, scoring scripts, configurations, example logs, analysis code, and trajectory- and step-level derivatives. Experiments show that episode traces expose process costs hidden by final-answer scoring, show how MDT-style consultation reallocates resources, and support longitudinal analyses of memory maturation, held-out transfer, update-stage response, and backward retention. Together, these results show that MedEvoEval provides a concrete basis for evaluating whether doctor agents improve through experience, transfer useful behavior, and retain earlier capabilities over time.
Multi-head latent attention (MLA) is increasingly important for long-context LLM inference because compact latent states replace the growing key-value (KV) cache and reduce decoding memory traffic. Yet most capable open checkpoints use multi-head or grouped-query attention (MHA/GQA), so conversion is needed to obtain MLA's cache efficiency without retraining from scratch. Speculative decoding offers complementary acceleration, but its speedup depends on agreement between draft proposals and target verification. We find that direct MHA/GQA-to-MLA conversion can sharply reduce this agreement: low-rank factorization and RoPE handling introduce attention-function errors that may be tolerable for standalone generation but substantially lower draft-token acceptance. We therefore formulate MLA draft construction as functional reconstruction rather than cache compression. Our end-to-end (E2E) method optimizes each converted MLA attention module to reproduce the post-output-projection response of its original MHA/GQA counterpart on calibration hidden states. This converter-agnostic post-conversion procedure preserves the converted cache and inference graph and requires neither verifier logits nor verifier supervision. We evaluate 192 model-converter-backend-method-task configurations spanning four Llama/Qwen draft-target pairs, TransMLA and MHA2MLA, HF and vLLM, and four 200-prompt tasks. With a 0.5-percentage-point reporting tolerance, Functional Reconstruction materially improves acceptance in 37 of 64 matched task cells, leaves 26 practically unchanged, and materially decreases one. Code and evaluation artifacts are available at https://github.com/swyhahaha/FunctionalMLA.
대규모 언어 모델(LLM) 기반 검색 에이전트(search agent)를 학습(training)시키려면 진정한 다중 홉(multi-hop) 검색을 요구하는 태스크와 검색 도구를 효과적으로 사용하는 궤적과 같은 고품질 검색 데이터가 필요합니다. 기존 파이프라인은 주로 사람이 작성한 태스크, 전문가 데모 또는 더 강력한 교사 모델(teacher model)에 의존합니다. 본 논문에서는 로컬 검색 환경(environment)에서 단일 LLM이 스스로 생성하고, 해결하며, 검증한 검색 태스크를 통해 해당 LLM을 학습시키는 자기 대국(self-play) 프레임워크인 SearchMaster를 제안합니다. 주요 과제는 자체 생성된 태스크와 롤아웃이 오해의 소지가 있는 신호를 생성할 수 있다는 점입니다. 즉, 가짜 다중 홉 질문, 검색 깊이를 무시한 성공률 기반 난이도 추정, 과도한 열람만 발생하고 목적 지향적인 증거 획득은 부족한 롤아웃 등이 발생할 수 있습니다. SearchMaster는 세 가지 제어 장치를 통해 이러한 실패 모드를 해결합니다. 첫째, 증거 사슬 생성기(Evidence-Chain Generator, ECG)는 명시적인 문서 간 증거 사슬에 태스크 생성을 그라운딩(grounding)하여 가짜 다중 홉 질문을 줄입니다. 둘째, 검색 깊이 보상(Search-Depth Reward, SDR)은 단순 성공률 대신 성공적인 롤아웃의 검색 깊이로 태스크 난이도를 평가하여, 유지되는 태스크가 검색 집약적이도록 보장합니다. 셋째, 과도 열람 페널티(Over-Opening Penalty, OOP)는 과도한 문서 열람을 억제하여 도구 사용을 규제함으로써 길지만 얕은 탐색을 방지합니다. 검증된 제안자(Proposer) 및 해결자(Solver) 롤아웃은 이후 GRPO를 통해 공동으로 최적화됩니다. 6개의 심층 검색 벤치마크(benchmark) 전반에 걸쳐 SearchMaster는 Qwen3.5-9B 백본의 평균 정확도를 38.19%에서 51.52%로 향상시켰으며, BrowseComp-Plus에서는 30.1%p의 향상을 달성했습니다. 이러한 결과는 근거 기반의 규제형 자기 대국이 사람이 라벨링한 QA 쌍이나 전문가 데모 없이도 효과적인 검색 에이전트 학습 데이터를 제공할 수 있음을 보여줍니다. 코드는 https://github.com/WentaoTan/SearchMaster 에서 이용할 수 있습니다.
학습과 테스트 이미지가 서로 다른 생성기(generator) 또는 데이터셋에서 유래할 때, AI 생성 이미지 탐지기는 일반화가 잘 되지 않는다. DINO와 같은 비전 기반 모델(vision foundation model)이 풍부한 공간 표현을 만들어냄에도 불구하고, 기존 탐지기는 대개 전역적으로 집약된 CLS 토큰만으로 이미지를 분류한다. 우리는 DINO 특징을 단일 CLS 토큰으로 전역 집약하면 공간적으로 분포된 생성 흔적이 가려진다고 가설을 세운다. 이 가설을 검증하기 위해 PatchHead를 제안한다. PatchHead는 DINO 패치 토큰의 2차원 구조를 유지하고 인접 영역에 걸친 증거를 통합하는 경량 공간 집약 헤드(spatial aggregation head)이다. 학습 동안 사전학습된 DINO 백본을 고정(freeze)하고, 삽입한 LoRA 어댑터, PatchHead, 보조 투영(projection) 헤드만 최적화한다. 수동 큐레이션과 in-the-wild 설정을 아우르는 9개의 교차 데이터셋 벤치마크에서 PatchHead는 7개 데이터셋에서 1위, 나머지 2개에서 2위를 기록한다. 기존 최강 방법 대비 평균 균형 정확도(balanced accuracy)를 91.6%에서 94.6%로 (+3.0 포인트) 끌어올리고, 최악 케이스 정확도를 82.4%에서 89.4%로 (+6.9 포인트) 향상시키면서, 학습 가능 파라미터는 8.6%만, FLOPs는 0.08%만 추가한다. 추가 정성 분석은 PatchHead가 (i) 클래스 조건부 도메인 불일치(class-conditional domain discrepancy)를 줄이고, (ii) 표현을 콘텐츠 지배적 saliency에서 공간적으로 분포된 진위(authenticity) 증거 쪽으로 전환시킨다고 시사한다. 이러한 관찰은 함께, 단일 CLS 기반 전역 표현보다 공간 패치 집약이 생성기와 데이터셋 간에 더 안정적으로 전이(transfer)되는 이유에 대한 표현 수준의 설명을 제공한다. 코드와 모델은 논문 게재(acceptance) 후 공개할 예정이다.