PyTorchKR Papers

최신 논문

총 523,801편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.00828v2cs.CV

Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA

Multi-camera driving question answering provides a discrete test of vision-language models, but post-training gains can reflect changed selected answers or improved answer extraction. We investigate this distinction within the Gemma model family using archived, task-linked WaymoQA evaluations. A zero-shot diagnostic on 5,519 test questions spans effective-small, mixture-of-experts, unified, and dense checkpoints, with serving differences disclosed rather than interpreted as architecture effects. The strongest controlled evidence is a matched-harness Gemma-4-12B base/GRPO pair: accuracy rises from 73.78% to 75.43%, a gain of 1.65 percentage points with a scene-cluster bootstrap interval of [+0.84, +2.41]. There are 215 improvements and 124 regressions; 213 improvements start from an extractable but incorrect base answer, and only two recover previously unextractable responses. Crossing correctness with extraction state shows why correct noncompliant outputs cannot be counted as prediction errors. Category counts reveal a critical safety asymmetry: routine perception gains (+5.93 pp) mask a regression in safety-critical hazard prediction (-1.82 pp), demonstrating that unweighted outcome rewards can trade off rare edge cases for aggregate accuracy. Reward saturation describes sampled training dynamics rather than an architectural ceiling. A separate archived Gemma-31B thinking comparison yields no detected average improvement, with uncertainty that does not establish equivalence. These results support a modest answer-selection gain for the recorded tuning configuration and a reproducible diagnostic accounting, without attributing corrected labels to specific grounding mechanisms or claiming superiority over matched supervised tuning.

arXiv:2611.00769v2cs.CV

DramaVQA: Probing Detailed Understanding of Ultra-Long TV Dramas

Despite recent progress in long-form video understanding, current benchmarks mostly focus on short clips or hour-scale videos. This leaves a major gap in evaluating ultra-long video understanding, where videos often last tens of hours. To bridge this gap, we introduce DramaVQA, a large-scale benchmark built upon long-form TV dramas. Containing approximately 20K question-answering (QA) pairs, the scale of DramaVQA exceeds previous long-video datasets, such as Video-MME and LVBench, by an order of magnitude. Our benchmark has two main features that make it stand out from existing datasets. First, we establish a scalable, LLM-assisted annotation pipeline that uses multimodal large language models (MLLMs) to generate complex questions, followed by strict human checking to ensure high data quality. Second, DramaVQA introduces a multi-level time structure; queries are designed across different video lengths, with the hardest questions requiring overall understanding across the entire drama. Extensive evaluations show that state-of-the-art MLLMs face major challenges when tested on these very long videos, highlighting the need for agentic AI algorithms capable of managing long-term memory. The project is available at https://github.com/xinyuehuo/DramaVQA

arXiv:2611.00661v2cs.AI

Designing Behavioral Profiles for Language Model Simulation of Out-of-Distribution Human Behavior

Large language models (LLMs) are increasingly used to simulate human behavior, but which descriptions of people improve prediction when experimental designs change remains unclear. We develop a framework that draws candidate behavioral features from the literature and constructs profiles from earlier human experiments. An LLM-based profile generator assigns feature values from observed actions, messages, and surveys; a separate LLM simulates behavior using these profiles. By comparing simulated and observed outcomes in the training data, we select profile contents through forward feature selection, adding features while they improve prediction. We then evaluate the selected contents without modification against outcomes from new participants under new designs. Across 616 public-goods games, 144 multi-party bargaining games, and 1,480 generalized 11--20 money-request designs, leading one-feature profiles predict human outcome distributions better than no profile or general-purpose personas. Profiles with more features provide no clear additional benefit, and features' relative predictive value largely persists across learning and validation. Using profiles from one game family to simulate another produces uneven gains and can make predictions less accurate than simulations without profiles. The framework makes the choice of behavioral information a testable part of simulation design, helping researchers assess which descriptions remain useful as experimental conditions change.

arXiv:2611.00633v2cs.AI

Evidence-grounded large language models guide perovskite solar cell design and optimization

Using large language models to interpret experimental evidence, design discriminating experiments and formulate well-supported mechanistic explanations remains a central challenge in materials research. Here we establish PSM-Bench, an 800-question benchmark developed by perovskite researchers to evaluate scientific reasoning in perovskite photovoltaics. We also introduce OpenPerov, an expert-informed language-model framework that learns from expert-calibrated evidence records reconstructed from 4,035 training articles; these records link experimental strategies, controls and observations to mechanistic interpretations. Structured-evidence adaptation outperformed conventional question-answer training, and OpenPerov Flash ranked first among 13 closed-book systems. OpenPerov Pro raised Scientific Mastery further, from 87.6 to 89.1, by revising selected passages with retrieved evidence. On questions derived from subsequent studies, both models also outscored expert-authored answers. In an experimental test of self-assembled monolayer design, OpenPerov selected a K analogue that outperformed the Na analogue preferred by experts. Devices incorporating the selected contact reached a power conversion efficiency of 27.05% and retained 96.74% of their initial efficiency after 1,000 h of dark thermal ageing at 85 $^\circ$C. These results show that reconstructed scientific reasoning provides an effective basis for building domain models that interpret experiments and inform materials design.

arXiv:2611.00006v2cs.NE

Cultural Evolution Drives the Emergence of Conventions in Neural Emergent Communication

Recent studies of language evolution with neural agents often use models where agents can either speak or listen, where there are only two agents in the population, or where agents can share internal representations. These choices prevent the study of how language emerges through cultural evolution, i.e. through the cumulative adaptation of individual behaviours to fit the communicative needs of the community. We tackle this issue with a new model that trains populations of bi-directional neural agents in a fully distributed fashion to solve a referential game. We demonstrate the effectiveness of this approach on multiple image datasets and show that it is able to scale to large populations. We also show that more conventional languages are easier to learn, providing a direct advantage for cultural transmission. Overall, this work provides a novel approach to neural emergent communication that opens up new directions for evolutionary linguistics.

arXiv:2610.42147v2cs.AI

Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention

Agentic workflows are increasingly shifting towards long-horizon tasks, such as open-ended scientific work, where models benefit from long context windows. However, at ultra-long context (above 1M tokens), full attention is prohibitively expensive. A promising approach is sparse attention, in which a dedicated indexer selects a subset of keys for the attention. Yet current methods of indexing either still have quadratic cost in context size or lose accuracy at ultra-long context scale. We address this with Latent Sparse Attention (LSA), an efficient attention method that maintains accuracy at ultra-long context. In LSA, keys are hierarchically clustered in a learned latent space, and queries score the hierarchy of clusters instead of scoring every key, making prefill, decode and training faster at long context. We propose two variants, Distilled-LSA (D-LSA), which can be applied to pre-trained models via distillation, and Native-LSA (N-LSA), which is trained natively on the next-token loss alone. On GLM-5.3-FLASH (QWEN-3.8-27B) at 10M tokens, D-LSA speeds up one attention layer by 22.7x (72.4x) in prefill and 4.6x (42.5x) in decode over its original sparse attention (full attention), while keeping similar accuracy across a variety of tasks. N-LSA matches the accuracy of full attention with a training speedup of 1.6x per layer at 128k tokens. LSA offers practical and efficient attention for models to benefit from ultra-long context.

arXiv:2610.42010v2cs.CL

TurnTurk-GPT: Text-Based Turn-Taking Prediction for Turkish

This paper presents TurnTurk-GPT, a text-based turn-taking predictor for Turkish obtained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective on synthetic and real dialogue, and evaluates it on real Turkish conversation against rule baselines, a zero shot Turkish language model and a widely used multilingual end-of-turn detector. Evaluation uses word-level decision points on unpunctuated, lower-cased text approximating the output of streaming speech recognisers, recording-disjoint and speaker-disjoint folds, nested threshold selection and paired cluster bootstrap tests. TurnTurk-GPT outperforms the multilingual detector under both recording-disjoint and speaker-disjoint evaluation. Its gain over a zero-shot Turkish language model is largest for speakers seen in training, while on unseen speakers it matches this strong baseline. Removing the suffixes of the turn-final word reduces AP by up to 44%, against 2% for a sentence-internal word, locating the Turkish end-of-turn cue in verb-final morphology. Synthetic dialogue alone yields the weakest trained model, and when a speaker pauses the zero-shot and multilingual baselines remain close to chance. The model and the evaluation protocol are publicly released.

arXiv:2610.41994v2cs.AI

DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention

Linear attention scales linearly with sequence length, but its finite-state memory can limit long-context retrieval. In gated linear recurrences, memory retention depends on state decay, which is shaped jointly by gate parameters A_log and dt_bias and by input activations. In controlled Gated DeltaNet (GDN) experiments, both parameters remain highly correlated with their initial values after training, highlighting the importance of initialization. Therefore, we introduce Dual-Axis Memory Initialization (DAMI), which jointly configures these parameters before training. DAMI assigns geometric timescales across heads for varied retrieval distances and a low-high-low input-sensitivity profile across layers. A small activation sample calibrates the GDN-family implementation to match both behavioral targets without adding trainable parameters. Across GDN, KDA, GDN2, Mamba-2, and RWKV-7, DAMI improves average RULER-style retrieval accuracy by 5.29 to 16.71 percentage points over the corresponding random-initialization baselines. Performance on language modeling, commonsense reasoning, and 2K real-world retrieval remains broadly comparable to those baselines. Our code is available at https://speedai.com/research/dami.

arXiv:2610.41970v2cs.AI

When Does Test-Time Search Help World Action Models? An Empirical Audit of Candidate Selection and Simulator Integrity

World-action models increasingly apply test-time scaling, spending extra inference computation on candidate selection and adaptive allocation to improve control. Such gains are credible only when candidates carry exploitable signal, nominal compute savings become real speedup, and evaluation stays reliable. We audit all three on Cosmos Policy and RoboCasa with a premise-first, pre-registered methodology. On clean data, a learned value selector is 1.66 pp worse than random; a hindsight oracle gains only 1.40 pp, with an episode-level interval spanning zero; and 82% of states show identical candidate outcomes, leaving little selection headroom. Token pruning cuts per-forward FLOPs by up to 47% but yields at most a measured 1.15$\times$ end-to-end speedup on the deployed path (1.25$\times$ on a decode-excluded path), short of a 1.5$\times$ target, and our adaptive-allocation experiment is inconclusive under a power gate. We further identify and fix a silent rendering corruption whose rate rises sharply along compute-heavy planning paths while leaving the official direct-policy result unaffected, and release a detector, fix, and scripts. These findings show why test-time scaling should be audited as a chain of premises (decision signal, realized efficiency, evaluationintegrity) rather than by success rate alone. Conclusions are scoped to the tested configuration.

arXiv:2610.41381v2cs.SD

Biologically Inspired Predictive Gating Signed-Dyadic Coding for Mandarin Keyword Spotting

In continuous spiking keyword spotting, temporally correlated acoustic frames can repeatedly trigger memory access and neural computation at a dense input interface, even when hidden activity is sparse. Predictive processing and sensory adaptation suggest retaining predictable input locally and responding selectively to deviations. Inspired by this mechanism, we propose Predictive Gating Signed-Dyadic Coding (PG). PG maintains a local prediction for each input channel and transmits an event only when the residual exceeds a learned threshold. Each event carries a sign and a finite power-of-two magnitude, and the receiver reconstructs it with fixed-point shifts and additions. In Mandarin keyword spotting, INPUT-PG reaches 90.00% accuracy while reducing mean SRAM reads by 33.87% and inference cycles by 32.77%. PG provides a hardware interface for research on low-power continuous speech processing and event-based sensory systems.

arXiv:2610.41351v2cs.LG

Quantile Propagation Planning for Budgeted Tail-Risk Evaluation

Risk-aware planning must identify rare, costly outcomes using only a limited number of model evaluations. Monte Carlo rollouts spend this budget on noisy tail estimates, while planning from the mean can miss the tail entirely. We introduce quantile propagation planning (QPP), which propagates a small, fixed set of disturbance quantiles through the dynamics and integrates the resulting cost curve to evaluate both mean and tail risk. The method is exact at the propagated levels, converges as the quantile grid is refined, and admits a one-sided conservatism guarantee under stated structural conditions. In controlled overload and inventory tasks, QPP reduces realized tail cost and constraint violations relative to mean-rollout planning. On overload, it remains competitive with sampled evaluators that use substantially more trajectories. These results position QPP as a transparent, budget-controlled approach to tail-risk evaluation.

arXiv:2610.41290v3cs.AI

SAT-ACT: Learning CDCL Branching from Outcome-Derived Action Preferences

Learning-based approaches have increasingly been explored for SAT solving, with particular attention to improving branching decisions in conflict-driven clause learning (CDCL) solvers. These approaches derive branching guidance either from static structural properties of SAT formulas or from dynamic signals obtained through solver feedback and execution traces. However, these forms of supervision provide limited evidence about state-dependent branching effects, while different outcome measures may favor different branching choices. To address these limitations, we propose SAT-ACT, an outcome-derived learning framework for CDCL branching. We design three complementary outcome comparison criteria to extract action preferences from independent branching action rollouts initiated at the same solver state. The criteria identify preferences supported by the outcomes and leave comparisons unresolved when the evidence is insufficient or conflicting, instead of forcing an ordering using a predefined scalar metric. A state-conditioned network learns branching priorities from the extracted preferences and uses them to guide CDCL search directly without online action rollouts. We further introduce non-dominated heuristic action supervision, which retains the solver-selected action as a training target only when it remains non-dominated under these preferences. Solver-level experiments show that SAT-ACT consistently reduces CDCL search costs on three SAT families and improves PAR-2 on the SAT Competition benchmarks.

arXiv:2610.41112v2cs.AI

Calibration of Large Model Quantization based on HiFloat4 Format

Quantization of Large Language Models (LLMs) has become increasingly important for reducing the computational costs and resource requirements for practical deployment. Growing interest in this area has motivated the community to move beyond 8-bit quantization towards 4-bit precision, leading to the development of customized 4-bit formats such as MXFP4 and NVFP4. Recently, HiFloat4 (HiF4) has emerged as a competitive alternative which adopts a Block Floating-Point (BFP) format with a three-level scaling hierarchy, thereby achieving a significantly larger dynamic range. While quantization can substantially reduce memory usage and accelerate inference, it frequently introduces accuracy degradation. Calibration is therefore essential for narrowing the performance gap between low-precision quantized models and their full-precision counterparts. Nevertheless, existing calibration methods are mainly designed and validated using integer-precision formats, such as INT4, lacking support for HiF4 due to its hierarchical BFP structure. In this paper, we investigate three calibration approaches, namely HiGPTQ, HiAutoRound, and HiFlatQuant, to further enhance model quantization based on the HiF4 format. Extensive experiment s demonstrate that all three proposed approaches improve inference accuracy compared with direct HiF4 conversion, reducing the gap between HiF4 and BF16 to less than 1% and outperforming the state-of-the-art NVFP4 format across various models and downstream tasks.

arXiv:2610.40946v2cs.AI

Opening the Box Changes the Cat: LLMs Detect Chain-of-Thought Monitoring - What now?

Chain-of-thought (CoT) reasoning allows large language models (LLMs) to generate intermediate reasoning traces in natural language while solving complex tasks. These traces can support safety monitoring by exposing potentially harmful plans before actions are executed. In this work, we investigate the conditions under which LLMs infer the presence of CoT monitoring from direct and indirect environmental cues and examine how this inference affects their subsequent behavior. Our observations suggest that oversight is an interactive process: while a monitor inspects a model's reasoning, traces left by monitoring interventions can allow the model to infer that its thoughts are being observed and to adapt its behavior accordingly. We show that even subtle environmental cues, such as timing-related signals from timeouts and operation durations, can reveal the presence of monitoring and trigger models to adapt their behavior to evade oversight. Across repeated episodes and interaction loops, we further observe that models can form statistical estimates of the monitor's effectiveness and use these estimates to guide attempts to evade detection. These findings highlight the need to evaluate CoT-based oversight under adaptive conditions, where the monitoring process itself may influence the behavior it seeks to observe.

arXiv:2610.40890v2cs.AI

ChatGPT as a Surveillance Assistant? An Exploratory Case Study of AI-Assisted Surveillance Risks

Large language models (LLMs) can generate detailed guidance, functional code, and persuasive text at scale. As one of the most widely used consumer language model products as of 2026, ChatGPT gives broad public access to these capabilities, motivating concerns about its potential use in surveillance-related tasks. We present an exploratory case study of what potentially harmful surveillance assistance looks like in ChatGPT's public web interface. We contribute a taxonomy of seven surveillance risk categories based on prior LLM risk research and reports of deployed surveillance systems, instantiated as 110 prompts, along with a manually reviewed analysis of ChatGPT's responses. Prompts were issued using two approaches: 55 jailbreak prompts drawing on five types of elicitation strategies, and 55 plain-language direct prompts, each run alone, and if unsuccessful, again with a short fixed prefix. We find that ChatGPT provides substantive surveillance guidance for 40 of the 55 jailbreak prompts and for 33 of the 55 direct prompts without any prefix. Compliant responses also rarely included policy mentions or suggestions for privacy-preserving alternatives. We manually review every response and find that even when ChatGPT states a refusal, it may still supply detailed instructions, working code, or architectural guidance alongside brief ethical caveats.

arXiv:2610.40654v2cs.LG

Propose-Filter-Release: Scaling Noise to Typical Sensitivity via Private Filtering

Differential privacy requires that algorithms be stable to the change of any record in their input, even to an extreme outlier. Private algorithms thus commonly introduce noise scaled to the worst-case possible influence of any single record. This work studies how to build differentially private algorithms that can effectively tailor their noise to the typical case rather than the worst case. We introduce the general \emph{propose-filter-release} (PFR) framework, which, by privately filtering its input, turns a mechanism that protects the addition or removal of a \emph{typical} record into one that is differentially private in the standard sense, on \emph{every} input. We show how to instantiate PFR with a novel filtering step to obtain a general meta-algorithm. Given only a concave friendliness score (which measures how typical a record is with respect to a dataset) and a release mechanism that is friendly-DP (that is, private conditioned on the changed record being typical of the remaining data), it guarantees differential privacy on sufficiently large datasets. The FriendlyCore [TCK+22] and Propose-Test-Release [DL09] algorithms can be viewed as instantiations of our general approach. We illustrate the generality of PFR by instantiating it for scalar statistics, principal subspace estimation with a dimension-independent sample size, the KL divergence of user histograms, the effective resistances in a resistor network, and treatment-rule selection.

arXiv:2610.40131v3cs.CV

VideoMemBench: How Well Do MLLMs Remember in Video Conversations?

Multimodal large language models (MLLMs) can now process hour-long videos, making extended multi-turn dialogue over a single video a natural mode of use. Sustaining such a dialogue requires more than answering each question in isolation: the model must retrieve and reuse information established earlier in the conversation, both what it has seen and what it has said, an ability we term in-context video memory. Existing video QA benchmarks leave this ability unmeasured: their questions are mutually independent, never reference one another, and each is answerable from the video alone, so nothing ever requires carrying information across turns. We introduce VideoMemBench, in which each sample pairs a single video with a long multi-turn dialogue, whose turns are connected by typed cross-turn dependencies: an Episodic Dependency refers back to a visual episode described in an earlier turn, probing memory of what the model has seen, while an Answer Dependency builds on the model's own earlier answer, probing memory of what the model has said. Every dependency is verified considering two properties: 1) necessity, meaning the question cannot be answered without its earlier turns, and 2) non-leakage, meaning the question does not restate the information from its earlier turns. To enforce these properties, we design a construction pipeline with three LLM roles: a planner, a generator, and a judger, with each dialogue reviewed by human experts. VideoMemBench comprises 326 videos with 2,476 turns, of which 1,387 depend on earlier turns, including multi-parent dependencies and long-range call-back. Among over 20 evaluated open- and closed-source MLLMs, Claude Opus 5 performs the best with an overall score of only 45.7%, highlighting the challenge of in-context video memory. Our dataset and evaluation are released.

arXiv:2610.39861v3cs.CV

M-PFT: Angular Margin for Calibrated Prompt-Tuning of Fine-Tuned Vision-Language Models

Prompt-tuning has emerged as a parameter-efficient technique for adapting large-scale vision-language models (VLMs) to downstream tasks while keeping the pre-trained weights frozen. However, prompt-tuned VLMs often suffer from a base-to-novel dual miscalibration problem, where models typically show underconfidence on base classes and overconfidence on novel classes. Current VLM calibration approaches address this dual miscalibration either separately or jointly, but lack a geometric understanding of the hyperspherical embedding space and thus fail to exploit angular discriminative properties to adjust distorted decision boundaries. To address this, we propose M-PFT, a novel angular-margin approach for prompt tuning that introduces two complementary objectives to promote class discriminability and semantic anchoring. This enlarges the angular margin by 1) minimizing the maximum cosine similarity between tunable class embeddings to reduce embedding collapse and 2) minimizing the angular distance between each tunable embedding and its corresponding zero-shot embedding to preserve semantic geometry. Extensive experiments across 9 prompt-tuning methods with various backbones on 15 diverse datasets show that our approach reduces the average calibration error on novel classes, while remaining competitive on base classes and maintaining comparable accuracy. Further analyses reveal that M-PFT effectively adjusts distorted decision boundaries induced by prompt tuning while preserving the pre-trained embedding space, improving calibration with negligible computational overhead.

arXiv:2610.39715v2cs.CV

NIF: Normal-Intersection Fields for Geometrically Consistent Oriented Box Refinement

Accurate oriented detection requires boxes to follow object boundaries, yet small errors in the coupled center, size, and angle estimates can displace long boundaries even when an object is correctly classified. We propose NIF, a plug-in refinement branch that represents box geometry with normal-intersection fields. Given a coarse oriented box, it samples local features at 20 boundary anchors and predicts signed distances along their outward normals. These observations have an exact affine representation in five rectangle parameters; a differentiable projection uses error-aware reliability weights to recover a geometrically consistent box. Tangential pooling preserves the normal-position profile, while the shared projection adds 0.132M parameters. On DOTA-v1.0, NIF obtains 77.98%, 79.03%, and 79.32% mAP with Oriented R-CNN, Strip R-CNN, and LSKNet-S, respectively; LSKNet-S gives the strongest single-scale result among the compared methods. NIF also reaches 68.19% mAP on DIOR-R and improves matched FAIR1M-v1.0 and DIOR-R baselines by 0.30% and 0.71%, respectively. On HRSC2016, it raises mean IoU from 0.867 to 0.888 while reducing field error by 24.5%. Finally, with four different detector frameworks frozen, NIF improves relative AP75 by 1.55-4.16% and COCO-style mAP by 0.37-3.76%, without detector retraining. Upon acceptance of this paper, we will publicly release the code at https://github.com/zsk466991125-collab/Normal-Intersection-Fields.

arXiv:2610.39688v2cs.CR

Hardware-Aware Model Inversion in Memristive Spiking Neural Networks

Hardware-accelerated spiking neural networks (SNNs) implemented on memristive crossbar arrays offer significant gains in energy efficiency, while device variability and analog noise are sometimes assumed to hinder information extraction. However, whether these hardware effects improve, preserve, or degrade privacy after deployment remains largely unexplored. Therefore, we ask whether model-inversion leakage characterized on an SNN remains representative after the same model is mapped to a non-ideal memristive substrate, and how hardware non-idealities and reliability faults affecting synapses, neuron thresholds, and spike processing reshape that leakage. To study this problem, we introduce Spiking Domain Model Inversion (SDMI), an inversion method that combines Expectation over Transformation (EOT) with surrogate gradients to optimize inputs through stochastic analog behavior and non-differentiable spiking dynamics. We use SDMI consistently across the SNN model, its memristive SNN, and the same memristive SNN under additional hardware perturbations. Across four datasets, reconstruction leakage does not vary monotonically with model accuracy or hardware variability. At comparable accuracy levels, neuron-threshold perturbations increase reconstruction fidelity, measured by SSIM, by up to 51.8%, whereas spike-level perturbations primarily degrade utility without consistently improving reconstruction. Weight noise, in contrast, reduces leakage under the same inversion procedure. These results show that hardware variability is not inherently privacy-preserving: model inversion leakage in memristive SNNs depends on the hardware state and cannot be inferred from model accuracy alone.