Inference-time techniques are emerging as highly effective tools to enhance large language model (LLM) capabilities. However, best practices for developing systems that combine these techniques remain underdeveloped due to our limited understanding of the utility of individual inference-time techniques and the interactions between them. Additionally, efficiently and automatically searching the space of model choices, inference-time techniques, and their compositions is challenging due to the large design space. To address these challenges, we introduce Archon, a modular framework for selecting, combining, and stacking layers of inference-time techniques to construct optimized LLM systems for target benchmarks. Rather than relying on a single LLM called once, we leverage a diverse set of LLMs and inference-time techniques, creating LLM systems greater than the sum of their parts. Archon defines an extensible design space, encompassing techniques such as generation ensembling, repeated sampling, ranking, fusion, critiquing, verification, and unit testing. It transforms the problem of building LLM systems into a hyperparameter optimization objective. Given the available LLMs, inference-time techniques, and compute budget, Archon utilizes hyperparameter search techniques to discover optimized architectures for target benchmark(s). We evaluate Archon architectures across a range of instruction-following, reasoning, and coding benchmarks, including MT-Bench, Arena-Hard-Auto, AlpacaEval 2.0, MixEval, MixEval Hard, MATH, and CodeContests. Archon architectures outperform frontier models, such as GPT-4o and Claude 3.5 Sonnet, on these benchmarks, achieving an average accuracy increase of 15.1 percentage points by using all available LLMs. We make our code and datasets available publicly on Github: https://github.com/ScalingIntelligence/Archon.
The utilization of large foundational models has a dilemma: while fine-tuning downstream tasks from them holds promise for making use of the well-generalized knowledge in practical applications, their open accessibility also poses threats of adverse usage. This paper, for the first time, explores the feasibility of adversarial attacking various downstream models fine-tuned from the segment anything model (SAM), by solely utilizing the information from the open-sourced SAM. In contrast to prevailing transfer-based adversarial attacks, we demonstrate the existence of adversarial dangers even without accessing the downstream task and dataset to train a similar surrogate model. To enhance the effectiveness of the adversarial attack towards models fine-tuned on unknown datasets, we propose a universal meta-initialization (UMI) algorithm to extract the intrinsic vulnerability inherent in the foundation model, which is then utilized as the prior knowledge to guide the generation of adversarial perturbations. Moreover, by formulating the gradient difference in the attacking process between the open-sourced SAM and its fine-tuned downstream models, we theoretically demonstrate that a deviation occurs in the adversarial update direction by directly maximizing the distance of encoded feature embeddings in the open-sourced SAM. Consequently, we propose a gradient robust loss that simulates the associated uncertainty with gradient-based noise augmentation to enhance the robustness of generated adversarial examples (AEs) towards this deviation, thus improving the transferability. Extensive experiments demonstrate the effectiveness of the proposed universal meta-initialized and gradient robust adversarial attack (UMI-GRAT) toward SAMs and their downstream models. Code is available at https://github.com/xiasong0501/GRAT.
This paper addresses the challenges of registering two rigid semantic scene graphs, an essential capability when an autonomous agent needs to register its map against a remote agent, or against a prior map. The hand-crafted descriptors in classical semantic-aided registration, or the ground-truth annotation reliance in learning-based scene graph registration, impede their application in practical real-world environments. To address the challenges, we design a scene graph network to encode multiple modalities of semantic nodes: open-set semantic feature, local topology with spatial awareness, and shape feature. These modalities are fused to create compact semantic node features. The matching layers then search for correspondences in a coarse-to-fine manner. In the back-end, we employ a robust pose estimator to decide transformation according to the correspondences. We manage to maintain a sparse and hierarchical scene representation. Our approach demands fewer GPU resources and fewer communication bandwidth in multi-agent tasks. Moreover, we design a new data generation approach using vision foundation models and a semantic mapping module to reconstruct semantic scene graphs. It differs significantly from previous works, which rely on ground-truth semantic annotations to generate data. We validate our method in a two-agent SLAM benchmark. It significantly outperforms the hand-crafted baseline in terms of registration success rate. Compared to visual loop closure networks, our method achieves a slightly higher registration recall while requiring only 52 KB of communication bandwidth for each query frame. Code available at: \href{http://github.com/HKUST-Aerial-Robotics/SG-Reg}{http://github.com/HKUST-Aerial-Robotics/SG-Reg}.
Micro-expressions (MEs) are crucial psychological responses with significant potential for affective computing. However, current automatic micro-expression recognition (MER) research primarily focuses on discrete emotion classification, neglecting a convincing analysis of the subtle dynamic movements and inherent emotional cues. The rapid progress in multimodal large language models (MLLMs), known for their strong multimodal comprehension and language generation abilities, offers new possibilities. MLLMs have shown success in various vision-language tasks, indicating their potential to understand MEs comprehensively, including both fine-grained motion patterns and underlying emotional semantics. Nevertheless, challenges remain due to the subtle intensity and short duration of MEs, as existing MLLMs are not designed to capture such delicate frame-level facial dynamics. In this paper, we propose a novel Micro-Expression Large Language Model (MELLM), which incorporates a subtle facial motion perception strategy with the strong inference capabilities of MLLMs, representing the first exploration of MLLMs in the domain of ME analysis. Specifically, to explicitly guide the MLLM toward motion-sensitive regions, we construct an interpretable motion-enhanced color map by fusing onset-apex optical flow dynamics with the corresponding grayscale onset frame as the model input. Additionally, specialized fine-tuning strategies are incorporated to further enhance the model's visual perception of MEs. Furthermore, we construct an instruction-description dataset based on Facial Action Coding System (FACS) annotations and emotion labels to train our MELLM. Comprehensive evaluations across multiple benchmark datasets demonstrate that our model exhibits superior robustness and generalization capabilities in ME understanding (MEU). Code is available at https://github.com/zyzhangUstc/MELLM.
Combinatorial problems are present in a wide range of industries. Constraint Programming (CP) is a well-suited problem-solving paradigm, but its core process, namely constraint modelling, is a bottleneck for wider adoption. Aiming to alleviate this bottleneck, recent studies have explored using Large Language Models (LLMs) as modelling assistants, transforming combinatorial problem descriptions to executable constraint models, similar to coding assistants. However, the existing evaluation datasets for constraint modelling are often limited to small, homogeneous, or domain-specific instances, which do not capture the diversity of real-world scenarios. This work addresses this gap by introducing CP-Bench, a novel benchmark dataset that includes a diverse set of well-known combinatorial problem classes sourced from the CP community, structured explicitly for evaluating LLM-driven CP modelling. With this dataset, and given the variety of constraint modelling frameworks, we compare and evaluate the modelling capabilities of LLMs for three distinct constraint modelling systems, which vary in abstraction level and underlying syntax: the high-level MiniZinc language and Python-based CPMpy library, and the lower-level Python interface of the OR-Tools CP-SAT solver. In order to enhance the ability of LLMs to produce valid constraint models, we systematically evaluate the use of prompt-based and inference-time compute methods adapted from existing LLM-based code generation research. Our results underscore the modelling convenience provided by Python-based frameworks, as well as the effectiveness of documentation-rich system prompts, which, augmented with repeated sampling and self-verification, achieve further improvements, reaching up to 70\% accuracy on this new, highly challenging benchmark.
강화학습(RL) 미세조정은 샘플링된 롤아웃에 대한 보상 신호를 사용하여 대규모 언어 모델(LLM)의 추론 능력을 향상시킨다. 보상 레이블이 롤아웃의 일부에 대해서만 존재할 때 정책 업데이트를 그 부분집합으로 제한하면 나머지 롤아웃은 정책 최적화에 사용되지 않는다. 준지도 학습은 레이블이 있는 샘플에서 없는 샘플로 레이블을 전파하지만, 기존 방법은 고정된 유사도 가중치로 레이블을 확산시키거나 각 샘플을 독립적으로 취급하여 질의, 사고 과정, 답변 사이의 관계를 무시한다. 본 연구는 온라인 정책 최적화 중 레이블이 없는 롤아웃으로 보상을 전파하는 그래프 기반 경험 메모리인 MemReward를 제안한다. MemReward는 초기 LLM 정책의 질의와 롤아웃(사고 과정과 최종 답변)을 유사도 및 구조적 에지로 연결된 이종 그래프의 노드로 저장하고, 레이블이 있는 롤아웃으로 GNN을 워밍업하여 이 그래프 위에서 보상을 전파한다. 온라인 RL 미세조정 동안 레이블이 없는 롤아웃은 질의 유사도에 따라 그래프에 연결되며, 정책은 레이블이 있는 롤아웃에는 정답 보상으로, 나머지에는 GNN이 예측한 보상으로 업데이트된다. 검증 가능한 보상이 있는 수학, 질의응답, 코드 생성의 13개 벤치마크에서 MemReward를 평가하고, 추가로 LLM 심사자, 인간 선호, 연속 보상으로도 평가한다. 롤아웃의 20%에만 정답 보상이 있을 때, MemReward는 Qwen2.5-1.5B에서 완전 지도 학습(Oracle) 성능의 96.6%, Qwen2.5-3B에서 97.3%를 달성하며, 이 경우 레이블 전파와 FixMatch 방식의 자기 학습도 능가한다. 또한 최대 8B까지의 Llama 백본으로도 확장되며 도메인 외 과제에서 Oracle에 근접한다.
대규모 언어 모델(LLM)의 추론 능력은 대부분 기반 모델에 잠재되어 있으며, 사후 학습(post-training)은 이를 도입하기보다 주로 증폭시킨다는 연구 결과가 축적되고 있다. 그러나 이러한 근거는 주로 수학 및 코딩 벤치마크에서 도출된 것으로, 해당 주장의 경계 조건, 즉 어떤 인지 과제가 유도(elicitation)를 통해 복원 가능하고 어디서 복원에 실패하는지는 아직 충분히 탐구되지 않았다. 이를 규명하기 위해 우리는 이론 기반 유도 도구인 NeuReasoner를 제안한다. 각 단계에서 오케스트레이터는 기능 특정성(functional specificity)에서 영감을 받은 Neuro Lens와 추론의 에로테틱 이론(Erotetic Theory of Reasoning)에서 도출한 Cognitive Lens를 결합하고, 외부 도구 없이 단일 모델의 내부 모듈화를 통해 두 출력을 통합한다. 우리는 NeuReasoner를 인지심리학의 행동 과제 모음인 CogBench와 표준 수학 및 코딩 벤치마크에서 평가하며, 기본 추론(vanilla inference) 대비 성능 향상과 모델의 사후 학습된 사고 모드(thinking mode)에 얼마나 근접하는지를 측정한다. 충분한 규모에서 NeuReasoner는 산술 추론, 코드 생성, 베이즈 추론, 보상 학습에서 사고 모드 기준선과 동등하거나 이를 초과하며, 이러한 성능 향상은 NeuReasoner의 결정당 호출 예산에 맞춘 자기 일관성(self-consistency) 및 반복 정제(iterative-refinement) 기준선에 대해서도 유지된다. NeuReasoner를 활용함으로써 명확한 경계를 발견할 수 있었다: 불확실성 하의 위험 감수 및 의사 결정은 유도만으로는 복원이 여전히 어렵고, 모델 규모는 유도와 양방향으로 상호작용하여 일부 인지 특성에서는 유도의 이점을 확대하고 다른 특성에서는 소멸시킨다. 전반적으로, NeuReasoner를 모듈식·해석 가능한 이론 기반 유도 도구로 활용하여, 기존 주장이 의존해온 수학 및 코딩 벤치마크를 넘어 추론 유도가 성공하고 실패하는 지점을 실증적으로 매핑한다.
추론 능력을 갖춘 대규모 언어 모델(LLM)을 정확도만으로 벤치마킹하면 핵심적인 배포 관련 질문, 즉 확장된 사고 토큰(thinking token)이 언제 그 비용만큼의 가치를 하는가라는 질문을 놓치게 된다. 본 논문은 추론 모델이 비추론 기준선(baseline) 대비 얻는 정확도 향상을 생성된 토큰 배수로 정규화하여 측정하는 한계(marginal) 벤치마킹 지표인 Token Economy Score(TES)를 제안한다. 추론 토글(toggle)을 가진 모델 패밀리와, 직접적인 비추론 대응 모델이 없는 최첨단(frontier) 모델을 위해 각각 쌍(paired) TES와 근사(approximated) TES 변형을 정의한다. 이어서 수학, 코드 생성, 과학 추론, 인스트럭션 팔로잉, 전문가 지식, 지식 회상, 연구 수준 물리학을 아우르는 7개 벤치마크에서 총 151회의 모델-벤치마크 평가 실행에 대한 실증적 벤치마킹 분석을 수행한다. 이 분석은 배포와 직결되는 세 가지 차원, 즉 어떤 과제 구조가 양의 한계 추론 효율을 내는지, 추론 노력(reasoning effort)을 높이면 모델 패밀리 내에서 TES가 어떻게 변하는지, 그리고 배포 환경에 따라 경제적 타당성이 어떻게 달라지는지를 살펴본다. 결과에 따르면 추론 효율은 명목상의 난이도보다 과제 구조가 더 잘 예측한다. AIME 2025와 LiveCodeBench처럼 순차적 추론 사슬(inference chain)이 필요한 과제는 높은 TES를 보인 반면, MMLU-Pro와 같은 지식 회상 과제는 어려움에도 불구하고 낮은 TES를 보였다. 또한 더 높은 추론 노력 수준에서는 수확 체감이 체계적으로 나타나며, 추가적인 사고가 오히려 정확도를 떨어뜨리는 경우도 있음을 확인했다. 마지막으로 추론 비용 비중(Reasoning Cost Share, RCS)은 추론(inference) 비용의 상당 부분이 내부 사고에 의해 좌우되는 경우가 많음을 보여주며, 배포 비용 배수(Deployment Cost Multiplier, DCM)는 온프레미스 배포가 그렇지 않으면 비용이 큰 추론 워크로드의 경제성을 어떻게 바꿀 수 있는지를 보여준다. 이러한 결과는 벤치마킹에 기반한 모델 선택 규칙, 즉 추론을 모든 경우에 유익한 모드로 취급하기보다 과제 유형, 노력 수준, 배포 환경에 따라 선택적으로 활성화해야 한다는 규칙을 뒷받침한다.
오픈 로봇 정책은 점차 두 가지 패러다임을 따르고 있습니다. 비전-언어-행동 모델(VLA)은 관측값과 지시사항을 행동으로 직접 매핑하는 반면, 월드-행동 모델(WAM)은 학습된 비디오 또는 월드 역학을 정책 학습이나 행동 생성에 통합합니다. 두 방식 모두 동일한 조작 작업을 목표로 하며 대안적인 설계 선택을 나타내지만, 일반적으로 서로 다른 평가 프로토콜 하에서 보고되어 능력, 강건성, 언어 민감도 및 배포 비용 간의 트레이드오프가 불명확한 상태로 남아 있습니다. 본 연구에서는 통합된 보고 체계 하에서 공개된 6개의 VLA 및 WAM 시스템에 대한 근거 인식(evidence-aware) 평가인 IndustrialVLA-Bench를 제시합니다. 이는 LIBERO에서의 기본 능력, LIBERO-Plus에서의 비언어적 강건성, LIBERO-Para에서의 지시사항 민감도, 관측된 실행 비용을 분리하여 평가합니다. 보고된 작업 점수는 고정된 체크포인트와 추론 구성 하에서 서로 다른 무작위 시드를 사용한 3회의 완전한 평가를 집계한 결과입니다. 6개 시스템 전체에서 기본 LIBERO 평균 점수 차이는 1.58점에 불과한 반면, 강건성 및 패러프레이즈 요약 점수의 차이는 각각 14.62점과 31.08점에 달합니다. 프로토콜을 충실히 따른 3개 시스템으로 비교를 한정하더라도 이러한 양상(1.36, 14.62, 23.10점)이 유지되므로, 여기서 보고된 진단적 분리는 더 낮은 근거 계층에 의존하지 않습니다. 또한 모든 시스템에 대해 관측된 추론 지연 시간, 최대 메모리, 런타임 모드 및 근거 상태를 함께 보고합니다. 프로토콜 준수, 준재현, 검증 대기 항목은 명확히 구분되어 유지되며, 프로토콜 준수 항목만이 엄격한 비교를 뒷받침합니다. 어느 한 패러다임의 보편적 우월성을 주장하기보다는, IndustrialVLA-Bench는 공유된 실용적 기준에 따라 공개된 로봇 정책을 비교할 수 있는 추적 가능한 근거를 제공합니다. 코드와 평가 기록은 https://github.com/xiaoqi-7/IndustrialVLA-Bench 에서 확인할 수 있습니다.
The latent space of a Generative Adversarial Network (GAN) has been shown to encode rich semantics within some subspaces. To identify these subspaces, researchers typically analyze the statistical information from a collection of synthesized data, and the identified subspaces tend to control image attributes globally (i.e., manipulating an attribute causes the change of an entire image). By contrast, this work introduces low-rank subspaces that enable more precise control of GAN generation. Concretely, given an arbitrary image and a region of interest (e.g., eyes of face images), we manage to relate the latent space to the image region with the Jacobian matrix and then use low-rank factorization to discover steerable latent subspaces. There are three distinguishable strengths of our approach that can be aptly called LowRankGAN. First, compared to analytic algorithms in prior work, our low-rank factorization of Jacobians is able to find the low-dimensional representation of attribute manifold, making image editing more precise and controllable. Second, low-rank factorization naturally yields a null space of attributes such that moving the latent code within it only affects the outer region of interest. Therefore, local image editing can be simply achieved by projecting an attribute vector into the null space without relying on a spatial mask as existing methods do. Third, our method can robustly work with a local region from one image for analysis yet well generalize to other images, making it much easy to use in practice. Extensive experiments on state-of-the-art GAN models (including StyleGAN2 and BigGAN) trained on various datasets demonstrate the effectiveness of our LowRankGAN.
With the development of machine learning techniques, the attention of research has been moved from single-modal learning to multi-modal learning, as real-world data exist in the form of different modalities. However, multi-modal models often carry more information than single-modal models and they are usually applied in sensitive scenarios, such as medical report generation or disease identification. Compared with the existing membership inference against machine learning classifiers, we focus on the problem that the input and output of the multi-modal models are in different modalities, such as image captioning. This work studies the privacy leakage of multi-modal models through the lens of membership inference attack, a process of determining whether a data record involves in the model training process or not. To achieve this, we propose Multi-modal Models Membership Inference (M^4I) with two attack methods to infer the membership status, named metric-based (MB) M^4I and feature-based (FB) M^4I, respectively. More specifically, MB M^4I adopts similarity metrics while attacking to infer target data membership. FB M^4I uses a pre-trained shadow multi-modal feature extractor to achieve the purpose of data inference attack by comparing the similarities from extracted input and output features. Extensive experimental results show that both attack methods can achieve strong performances. Respectively, 72.5% and 94.83% of attack success rates on average can be obtained under unrestricted scenarios. Moreover, we evaluate multiple defense mechanisms against our attacks. The source code of M^4I attacks is publicly available at https://github.com/MultimodalMI/Multimodal-membership-inference.git.
Multi-hop question answering is a knowledge-intensive complex problem. Large Language Models (LLMs) use their Chain of Thoughts (CoT) capability to reason complex problems step by step, and retrieval-augmentation can effectively alleviate factual errors caused by outdated and unknown knowledge in LLMs. Recent works have introduced retrieval-augmentation in the CoT reasoning to solve multi-hop question answering. However, these chain methods have the following problems: 1) Retrieved irrelevant paragraphs may mislead the reasoning; 2) An error in the chain structure may lead to a cascade of errors. In this paper, we propose a dynamic retrieval framework called Tree of Reviews (ToR), where the root node is the question, and the other nodes are paragraphs from retrieval, extending different reasoning paths from the root node to other nodes. Our framework dynamically decides to initiate a new search, reject, or accept based on the paragraphs on the reasoning paths. Compared to related work, we introduce a tree structure to handle each retrieved paragraph separately, alleviating the misleading effect of irrelevant paragraphs on the reasoning path; the diversity of reasoning path extension reduces the impact of a single reasoning error on the whole. We conducted experiments on three different multi-hop question answering datasets. The results show that compared to the baseline methods, ToR achieves state-of-the-art performance in both retrieval and response generation. In addition, we propose two tree-based search optimization strategies, pruning and effective expansion, to reduce time overhead and increase the diversity of path extension. We will release our code.
Haze severely degrades the visual quality of remote sensing images and hampers the performance of road extraction, vehicle detection, and traffic flow monitoring. The emerging denoising diffusion probabilistic model (DDPM) exhibits the significant potential for dense haze removal with its strong generation ability. Since remote sensing images contain extensive small-scale texture structures, it is important to effectively restore image details from hazy images. However, current wisdom of DDPM fails to preserve image details and color fidelity well, limiting its dehazing capacity for remote sensing images. In this paper, we propose a novel unified Fourier-aware diffusion model for remote sensing image dehazing, termed RSHazeDiff. From a new perspective, RSHazeDiff explores the conditional DDPM to improve image quality in dense hazy scenarios, and it makes three key contributions. First, RSHazeDiff refines the training phase of diffusion process by performing noise estimation and reconstruction constraints in a coarse-to-fine fashion. Thus, it remedies the unpleasing results caused by the simple noise estimation constraint in DDPM. Second, by taking the frequency information as important prior knowledge during iterative sampling steps, RSHazeDiff can preserve more texture details and color fidelity in dehazed images. Third, we design a global compensated learning module to utilize the Fourier transform to capture the global dependency features of input images, which can effectively mitigate the effects of boundary artifacts when processing fixed-size patches. Experiments on both synthetic and real-world benchmarks validate the favorable performance of RSHazeDiff over state-of-the-art methods. Source code will be released at https://github.com/jm-xiong/RSHazeDiff.
As the rapid development of computer vision and the emergence of powerful network backbones and architectures, the application of deep learning in medical imaging has become increasingly significant. Unlike natural images, medical images lack huge volumes of data but feature more modalities, making it difficult to train a general model that has satisfactory performance across various datasets. In practice, practitioners often suffer from manually creating and testing models combining independent backbones and architectures, which is a laborious and time-consuming process. We propose Flemme, a FLExible and Modular learning platform for MEdical images. Our platform separates encoders from the model architectures so that different models can be constructed via various combinations of supported encoders and architectures. We construct encoders using building blocks based on convolution, transformer, and state-space model (SSM) to process both 2D and 3D image patches. A base architecture is implemented following an encoder-decoder style, with several derived architectures for image segmentation, reconstruction, and generation tasks. In addition, we propose a general hierarchical architecture incorporating a pyramid loss to optimize and fuse vertical features. Experiments demonstrate that this simple design leads to an average improvement of 5.60% in Dice score and 7.81% in mean interaction of units (mIoU) for segmentation models, as well as an enhancement of 5.57% in peak signal-to-noise ratio (PSNR) and 8.22% in structural similarity (SSIM) for reconstruction models. We further utilize Flemme as an analytical tool to assess the effectiveness and efficiency of various encoders across different tasks. Code is available at https://github.com/wlsdzyzl/flemme.
Large Language Models (LLMs) frequently lack domain-specific knowledge and even fine-tuned models tend to hallucinate. Hence, more reliable models that can include external knowledge are needed. We present a pipeline, 4StepFocus, and specifically a preprocessing step, that can substantially improve the answers of LLMs. This is achieved by providing guided access to external knowledge making use of the model's ability to capture relational context and conduct rudimentary reasoning by themselves. The method narrows down potentially correct answers by triplets-based searches in a semi-structured knowledge base in a direct, traceable fashion, before switching to latent representations for ranking those candidates based on unstructured data. This distinguishes it from related methods that are purely based on latent representations. 4StepFocus consists of the steps: 1) Triplet generation for extraction of relational data by an LLM, 2) substitution of variables in those triplets to narrow down answer candidates employing a knowledge graph, 3) sorting remaining candidates with a vector similarity search involving associated non-structured data, 4) reranking the best candidates by the LLM with background data provided. Experiments on a medical, a product recommendation, and an academic paper search test set demonstrate that this approach is indeed a powerful augmentation. It not only adds relevant traceable background information from information retrieval, but also improves performance considerably in comparison to state-of-the-art methods. This paper presents a novel, largely unexplored direction and therefore provides a wide range of future work opportunities. Used source code is available at https://github.com/kramerlab/4StepFocus.
With the advancement of large language models (LLMs), solving complex reasoning tasks has gained increasing attention. Inference-time computation methods (e.g., Best-of-N, beam search, et al.) are particularly valuable as they can enhance reasoning performance without modifying model parameters or requiring additional training. However, these techniques come with implementation challenges, and most existing methods remain at the proof-of-concept stage with limited practical adoption due to their computational complexity and varying effectiveness across different tasks. In this paper, we investigate and benchmark diverse inference-time computation strategies across reasoning tasks of varying complexity. Since most current methods rely on a proposer-verifier pipeline that first generates candidate solutions (e.g., reasoning solutions) and then selects the best one based on reward signals (e.g., RLHF rewards, process rewards), our research focuses on optimizing both candidate solution generation (e.g., instructing prompts, hyperparameters such as temperature and top-p) and reward mechanisms (e.g., self-evaluation, reward types). Through extensive experiments (more than 20,000 A100-80G GPU hours with over 1,000 experiments) across a variety of models (e.g., Llama, Qwen, and Mistral families) of various sizes, our ablation studies reveal that previously overlooked strategies can significantly enhance performance (e.g., tuning temperature can improve reasoning task performance by up to 5%). Furthermore, we establish a standardized benchmark for inference-time computation by systematically evaluating six representative methods across eight reasoning tasks. These findings provide a stronger foundation for future research. The code is available at https://github.com/usail-hkust/benchmark_inference_time_computation_LL
Recent advances in Generative AI (GenAI) have led to significant improvements in the quality of generated visual content. As AI-generated visual content becomes increasingly indistinguishable from real content, the challenge of detecting the generated content becomes critical in combating misinformation, ensuring privacy, and preventing security threats. Although there has been substantial progress in detecting AI-generated images, current methods for video detection are largely focused on deepfakes, which primarily involve human faces. However, the field of video generation has advanced beyond DeepFakes, creating an urgent need for methods capable of detecting AI-generated videos with generic content. To address this gap, we propose a novel approach that leverages pre-trained visual models to distinguish between real and generated videos. The features extracted from these pre-trained models, which have been trained on extensive real visual content, contain inherent signals that can help distinguish real from generated videos. Using these extracted features, we achieve high detection performance without requiring additional model training, and we further improve performance by training a simple linear classification layer on top of the extracted features. We validated our method on a dataset we compiled (VID-AID), which includes around 10,000 AI-generated videos produced by 9 different text-to-video models, along with 4,000 real videos, totaling over 7 hours of video content. Our evaluation shows that our approach achieves high detection accuracy, above 90% on average, underscoring its effectiveness. Upon acceptance, we plan to publicly release the code, the pre-trained models, and our dataset to support ongoing research in this critical area.
테스트 시간 스케일링(test-time scaling)은 코딩에서의 컴파일러 및 테스트 케이스나 로보틱스 응용에서의 학습된 가치 함수(value function)와 같은 외부 검증기(verifier)를 사용하여 고품질 롤아웃을 얻는 경우가 많습니다. 검증기 없는 테스트 시간 스케일링(VF-TTS)은 많은 실제 응용 분야에서 이러한 고품질 검증기에 접근할 수 없다는 점 때문에 대규모 언어 모델(LLM) 추론을 향상하는 메커니즘으로서 광범위한 주목을 받고 있습니다. 기존 VF-TTS 방법 중에서도 신뢰도(confidence)만을 기반으로 롤아웃을 계산하고 순위를 매기는 신뢰도 기반 VF-TTS 방법이 특히 유망합니다. 이러한 방법은 샘플 평가를 위한 오버헤드가 거의 없으며 모델 내부 상태에 대한 최소한의 접근만을 필요로 하므로, 다양한 모델과 작업 전반에서 매우 뛰어난 유연성을 제공합니다. 본 논문에서는 기존의 신뢰도 기반 VF-TTS 방법이 복잡한 작업에서 치명적으로 실패함을 보여줌으로써 이들의 결정적인 한계를 입증합니다. 우리는 매우 흥미로운 현상을 관찰했습니다. 바로 균일하게 높은 신뢰도는 탐색의 실패를 나타내는 경우가 빈번하며, 확신에 찬 오답을 선호하게 만든다는 점입니다. 이를 해결하기 위한 우리의 핵심 통찰은 강건한 인지 탐색에 특정한 신뢰도 궤적 패턴이 필요하다는 것입니다. 즉, 이러한 방법은 초기의 낮은 신뢰도로 나타나는 탐색적 분기를 초반에 수행하고, 최종적으로 높은 신뢰도를 갖는 해로 수렴해야 합니다. 이러한 통찰을 구현하기 위해, 우리는 추론에서 신뢰도의 시간적 비대칭성을 명시적으로 평가하는 새로운 선택 프레임워크인 통섭(consilience)을 도입합니다. 우리는 최종적인 확실성을 엄격히 요구하면서 초기의 높은 신뢰도에는 능동적으로 페널티를 부여하는 조합론적 지표를 통해 이를 구체화합니다. 대학원 수준의 수학 문제와 자유 형식 코드 생성을 모두 다루는 광범위한 실험을 통해 통섭이 기존 베이스라인을 효과적으로 능가함을 입증하고, 완결 신뢰도(completion confidence)에 대한 우리의 새로운 관점을 검증합니다.
Real-time spoken dialogue systems face a fundamental tension between latency and response quality. End-to-end speech-to-speech (S2S) models respond immediately and naturally handle turn-taking, backchanneling, and interruption, but produce semantically weaker outputs. Cascaded pipelines (ASR -> LLM) deliver stronger responses at the cost of latency that grows with model size. We present RelayS2S, a hybrid architecture that runs two paths in parallel upon turn detection. The fast path - a duplex S2S model - speculatively drafts a short response prefix that is streamed immediately to TTS for low-latency response onset, while continuing to monitor live audio events. The slow path - a cascaded ASR -> LLM pipeline - generates a higher-quality continuation conditioned on the committed prefix, producing an uninterrupted utterance. A lightweight learned verifier gates the handoff, committing the prefix when appropriate or falling back gracefully to the cascaded pipeline. With GPT-4.1 as the back-end, RelayS2S substantially reduces response latency while preserving nearly all of the cascaded pipeline's textual quality. On synthetic voice dialogues, it achieves a P90 first-chunk latency of 81 ms, excluding TTS and network latency, compared with 1,006 ms for the cascaded baseline. On real voice dialogues, RelayS2S reduces average first-chunk latency by 479 ms while retaining 99% of the cascaded pipeline's textual quality. These benefits become larger as the slow-path model scales. Because the prefix handoff requires no architectural modification to either component, RelayS2S serves as a lightweight, drop-in addition to existing cascaded pipelines. Our code is publicly available at: https://github.com/mailong25/relays2s
텍스트-투-이미지 확산 모델의 인상적인 생성 능력에도 불구하고, 이들은 암시적 성적 프롬프트에 취약하다. 미묘한 단서가 무해한 용어 또는 적대적 토큰으로 위장되어 모델 편향이나 학습 데이터의 잠재적 상관관계로 인해 부적절한 콘텐츠가 예기치 않게 생성된다. 기존 안전성(safety) 메커니즘은 근본적인 한계에 직면한다. 탐지 방법은 주로 노골적인 콘텐츠를 식별하고 암시적 악의적 의도를 포착하지 못하며, 완화 접근법은 다양한 암시적 시나리오에 부적합한 정적 부정 프롬프트(negative prompt)에 의존한다. 이러한 과제를 해결하기 위해, 우리는 확산 과정의 노이즈 역학 관점에서 안전성 메커니즘을 재고하는 통합 노이즈 기반 프레임워크인 UniNDM을 제안한다. 핵심 통찰은 초기 단계 예측 노이즈가 정상 콘텐츠와 성적으로 노골적인 콘텐츠 사이에서 고유한 분리 가능성을 보인다는 것이며, 이를 이론적으로 타임스텝(timestep)에 따라 이차적으로 증가하는 의미적 집중을 입증한다. 이 특성을 활용하여, 사실상 계산 오버헤드 없이 우수한 정확도를 달성하는 경량 노이즈 기반 탐지기를 개발한다. 완화를 위해 노이즈 강화 적응형 부정 가이던스(noise-enhanced adaptive negative guidance)를 도입한다. 대규모 언어 모델(LLM)을 통해 다양한 암시적 콘텐츠를 처리하기 위해 문맥 특화 부정 프롬프트를 동적으로 생성하고, 동시에 노골적인 토큰에 대한 어텐션 집중을 억제하여 초기 노이즈를 최적화함으로써 포괄적인 보호를 제공한다. U-Net 기반 확산 모델 외에도, 통합 멀티모달 어텐션에 맞춘 영역 제약 의미적 가이던스(region-constrained semantic guidance)를 통해 새롭게 부상하는 Diffusion Transformer(DiT) 아키텍처로 프레임워크를 확장한다. U-Net 모델과 DiT 모델에 걸친 자연 및 적대적 데이터셋에서의 포괄적 실험은 SLD, UCE, Safree 등 최신(state-of-the-art) 방법 대비 상당한 개선을 보여준다. 코드는 https://github.com/Aries-iai/UniNDM 에서 공개적으로 이용 가능하다.