PyTorchKR Papers

최신 논문

총 569,556편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.11209v1cs.CV

Viscot-steer: Disentangling and Transferring Visual Chain-of-Thought via Activation Steering

Visual chain-of-thought (visual-CoT), also known as "thinking with images", equips LVLMs with the ability to locate and use key visual evidence for reasoning. However, acquiring such visual-CoT capabilities requires costly post-training and tightly couples them to a specific backbone, making transfer to other LVLMs difficult. Existing cross-model transfer methods typically treat visual-CoT as a monolithic capability, using a unified representation that ignores the functional distinction between grounding and reasoning and introduces stage-irrelevant interventions. To address this problem, we propose Viscot-steer, a training-free activation steering framework for transferring visual-CoT capabilities across heterogeneous LVLMs. Specifically, Viscot-steer first disentangles visual-CoT into grounding and reasoning vectors and extracts them from the source-model layers associated with the corresponding capabilities. It then aligns the vectors with the activation space of the target model and steers the grounding and reasoning layers during bounding-box prediction and reasoning-trace generation, respectively. Across benchmarks covering high-resolution visual understanding, general visual perception and reasoning, and mathematical visual reasoning, Viscot-steer transfers capabilities from different visual-CoT models to architecturally diverse LVLMs and outperforms the strongest capability-transfer baseline by an average of 1.86 percentage points on the primary evaluation metrics.

arXiv:2611.11204v1cs.LG

Not Yet in the Profile: Uncertainty-Aware Dynamic User Profiling by Learning from Future Interactions

On the Web, user profiling builds compact user representations from interaction histories for efficient personalization, and dynamic user profiling updates them as users' behavior and preferences change. However, new user behavior can be ambiguous, and its interpretation differs across users. To address this, we introduce Uncertainty-aware Adaptive Profile Manager (UAPM), which separates the claims in a user profile from interpretations of new behavior that are not yet confirmed, through a Profile, a Buffer, and a Rulebook. UAPM records user-specific interpretation and update rules in the Rulebook, keeps ambiguous evidence in the Buffer outside the Profile as evidence for later decisions, and writes supported general and situational characteristics to the Profile. We train UAPM with Privileged On-Policy Distillation with Dual Specialization (POPD-DS), which learns from recorded future interactions during training without access to them at deployment. Its first stage, Dual Specialization, distills a stronger model into a current-only Student and a future-informed model that becomes the Teacher. In the second stage, privileged on-policy distillation (POPD), only the fixed Teacher receives future interactions and guides the Student's own management decisions alongside task rewards. Experiments on three Web personalization tasks show that UAPM outperforms multiple baselines under the same downstream input budget, improving on the best same-budget baseline by up to 5%.

arXiv:2611.11201v1cs.AI

When Self-Consistency Hurts: The Hidden Cost of Majority Voting in Small Language Models

We study prompted chain-of-thought (CoT) with self-consistency (SC), i.e. sampling N traces and plurality-voting, on small language models, a regime where per-query compute budgets are tight yet SC is commonly applied with a fixed sample budget inherited from large-model practice, regardless of instance difficulty. Across two model families (Qwen-2.5-0.5B, Gemma-3-1B) at FP16 and MLX-Q4 (4-bit quantisation in Apple's MLX framework), on GSM8K and MATH-500, we run 8 per-task and 4 task-pooled experiments at n=200-400 with 1,000x paired-bootstrap confidence intervals (CIs). Our first result is an upper bound: an oracle gate deciding per prompt whether to invoke SC reaches 109-174% of always-SC's accuracy at 23-30% of its compute in all 12/12 cases, so at N=5 always-SC is dominated everywhere on our grid. That margin is exactly the rate at which SC flips correct greedy answers to wrong, here 3-10.5% of prompts, which on Qwen-Q4 x MATH-500 exceeds the wrong-to-right rate and leaves always-SC 3 pt worse than always-greedy. Our second result is a deployable approximation: a gate over greedy-pass signals skips 45-99% of SC calls, with bootstrap CI half-widths under 5 pt, removing 36-79% of always-SC compute. Accuracy retention at those operating points has point estimates of 98.1-98.8%, which we report as point estimates because their CIs are wide (lower bounds 76-86%) over an SC accuracy of only 15-36%. The evidence covers N=5 on two <=1B families and two math benchmarks, which is the scope we claim. Code is available at https://github.com/heywanrong/SiftSC.

arXiv:2611.11200v1cs.CV

Source-Anchored Visual Autoregression for Extreme Image Compression

In the extreme-low bitrate range below $0.01$ bpp, image compression becomes a constrained generation problem: the reconstruction depends on inferring plausible missing details while staying close to the input image. Autoregressive token models provide an efficient coarse-to-fine prior for this setting, but the transmitted early-scale tokens often leave object identity and texture underdetermined, which can lead to source-inconsistent reconstructions. We introduce Source-Anchored Visual Autoregression (SAVAR), a generative codec built on Visual Autoregressive (VAR) modeling. SAVAR transmits early-scale tokens and a compact caption. The tokens provide image-specific layout information, while the caption supplies global semantics. SAVAR reconstructs a partial VQ-VAE latent from the retained scales, maps it to the start-of-sequence initialization, and uses the same latent to condition the caption embedding before adaptive layer normalization, separating two roles that are typically driven by a shared condition in VAR-style generation. On the Kodak and CLIC2020 datasets in the extreme-low range, SAVAR reduces perceptual distortion relative to matched ARPC retained-scale operating points, lowering LPIPS by about $8$--$9\%$ and DISTS by up to $22\%$ at a lower bpp. It reconstructs a $512 \times 512$ image in about $218$ ms on one NVIDIA L20X GPU, retaining fast autoregressive decoding in the evaluated setting.

arXiv:2611.11197v1cs.IR

Dhara: Evidence-Oriented Retrieval for Indian Knowledge Systems

Relevant evidence in Indian Knowledge Systems (IKS) collections may be distributed across passages in different parts of a work, including verses, translations, and commentaries. Locating it requires matching natural-language questions and partially remembered verse fragments to source passages. We present Dhara, a retrieval engine combining hybrid search, chapter- and book-level Entity Notes (EN), Fuzzy Verse Retrieval (FVR), and neural reranking. Five comparisons on the 325 questions in Dhara Bench examine retrieval representations against BM25, EN's contribution, reranker quality and computation time, FVR, and Gemini. BGE-M3 single-vector retrieval exceeds BM25 in nDCG@10, Recall@20, and MRR@20 across all categories. Its scores are broadly comparable to ColBERTv2, with mean retrieval times of 0.108 s and 0.962 s under the same hardware configuration and a smaller memory footprint for chunk storage on disk. Adding EN to a direct branch with a matched candidate count improves nDCG@10 in all twelve categories with Qwen and ten with BGE; the combined ranking is capped at the same candidate count. Qwen improves nDCG@10, Precision@5, and MRR@20 over BGE in every category, at approximately 6.72 times the overall estimated serial computation time. FVR increases Verse-category nDCG@10 by approximately 20 percentage points with either reranker. Gemini is evaluated after source mapping and manual review with relevance annotations visible. The Qwen pipeline with EN and FVR has higher Recall@20 than Gemini in ten categories and stronger complete-evidence retrieval in Multi-hop questions, while Gemini retains early-relevance advantages in several categories. The results inform retrieval choices according to evidence requirements and resource budgets.

arXiv:2611.11189v1cs.LG

TEAM: Topology-Aware, Semantic-Sensitive, General Transformer for Multimodal-Attributed Graphs

Multimodal-Attributed Graphs (MAGs) organize heterogeneous data in graphs whose nodes are characterized by modalities such as text and images rather than traditional unimodal attributes. However, many representative MAG methods organize propagation around node-level states and topology-defined neighborhoods, which can obscure fine-grained modality contributions: a bounded neighborhood may contain semantically weak evidence, while compatible context need not be confined to it. To address both limitations, we propose TEAM, a modality-centric framework that constructs and encodes graph context with node-modality tokens rather than treating each node as a single interaction unit. Concretely, TEAM unfolds the local topology into a Modality Encoding Tree and performs tree-anchored semantic evidence acquisition, organizing local structural context while expanding the accessible evidence space outside the constructed local Tree. We further introduce the Modality-Source Context Transformer (MSCT), which encodes the selected node-modality tokens and coordinates interactions among modality-source groups. By retaining textual and visual attributes as separate interaction units until readout, it supports fine-grained cross-node cross-modal exchange without first collapsing them into whole-node features. Across the evaluated baselines, TEAM attains the highest reported mean on all graph-learning metrics, all three G2Text metrics, and both G2Image metrics. On Toys, it surpasses the strongest baseline by 3.37 and 2.93 percentage points in Accuracy and Macro-F1, respectively.

arXiv:2611.11179v1cs.SD

Guidance at Pure Noise Erodes Speaker Similarity in Few-Step Token-Conditioned Flow-Matching TTS

Text-to-speech systems built from a language model and a flow-matching mel decoder now run the decoder with two to four Euler steps instead of the official ten or more. At two steps the word error rate (WER) is unchanged, but speaker similarity (SIM) drops by one to six points. This paper shows that most of this loss comes from classifier-free guidance on the first step. At pure noise the network can only return the conditional mean, so guidance shifts every trajectory by nearly the same vector. Later steps do not remove this shift, and its size grows with the length of the first step, so it is negligible with ten or more steps and accounts for most of the loss with two. Two training-free samplers are derived from this analysis. First-step unguided sampling (FSU) leaves the first step unguided and recovers $84$--$91\%$ of the two-step SIM loss on CosyVoice~2/3 with no significant WER increase. FSU-Mid replaces the two Euler steps by one midpoint step with an unguided first evaluation, recovers $77$--$85\%$ on IndexTTS-2 and matches the ten-step teacher on CosyVoice~2 and GLM-TTS. Both use three network function evaluations, one fewer than the official two-step sampler, and FSU-Mid gives the highest mean recovery among eight published guidance rules re-run at two steps on four decoders.

arXiv:2611.11177v1cs.AI

When Are Nearby Experts Composable? Cross-Task Model Merging in Neural Thickets

Random perturbations around a shared checkpoint can produce experts for distinct tasks, but shared origin does not ensure that their gains survive composition. We formulate expert-pair selection for joint base-relative retention, measured by strict joint success and worst-task gain with unconditional abstention accounting. An exact finite-difference decomposition and an arbitrarily small orthogonal counterexample show why proximity and orthogonality alone are insufficient. We instantiate this formulation with quality-constrained spectral selection, a fixed-budget framework combining cross-task screening, layerwise pair geometry, separate validation, and Base fallback. Across six checkpoints, 21 task pairs, and five independent perturbation pools, the procedure improves strict joint success over Top-1 averaging throughout and leads the four-method model-scale comparison on four of six checkpoints. For the five models with at least 1.5B parameters, matched quality-only, global-cosine, and spectral ranking attain 49.1%, 50.5%, and 53.1% joint success. Under this matched protocol, the spectral selector also yields the highest worst-task gain and Weak Pareto rate and the lowest regression rate. Mean worst-task gain reaches +0.61 percentage points, while 17.1% regression and 13.0% abstention show that composability remains conditional. Project page: https://anonymous.4open.science/r/Thickets-Merge/.

arXiv:2611.11174v1cs.DL

What Changed in Engineering Writing After Generative AI Chat Assistants: Evidence from 1.8 Million Articles

This study asks what changed in engineering articles after conversational language models became publicly available on November 30, 2022, and develops methods for measuring such change in a field without a uniform abstract index. It assembled 1,839,962 articles published from 2015 to 2025 in 256 engineering journals, tracked twelve measures of wording, style, and article structure, and compared each with the projection of its own pre-2023 trend. Building the monthly series showed that records dated only by year fall on January 1 and create false January peaks unless removed. In abstracts, 31 words associated with language model output appeared 82% more often than projected (55.4 against 27.0 per 10,000 words in 2025), while four control words of the same formal register fell 16%. Words that soften a claim, such as may, fell 13%, words that strengthen one, such as clearly, rose 20%, and long words rose 11%. Author counts, reference lists, page counts, and publication volume stayed within 4% of trend. The change appears in all fourteen subject categories, and false release dates, dropping any one journal or word, and word lists chosen from the data support it. A random sample of 12,866 full texts from the same corpus, including subscription articles and weighted back to it, shows the same words rising 99% in article bodies. Statements disclosing AI writing assistance, measured apart from mentions of language models, went from none to 3.7% in 2025, four in five of them at one publisher. The design is descriptive, not causal.

arXiv:2611.11170v1cs.LG

ClusLoad: Cluster-Guided Model Sharing for Scalable Distribution Transformer Load Forecasting

Large-scale device-level load forecasting for distribution transformers is challenged by substantial cross-device heterogeneity and the deployment burden of maintaining device-specific forecasting models. Existing approaches typically rely either on device-specific models, which are difficult to scale to large transformer populations, or on global models, which reduce deployment complexity but may overlook heterogeneous load patterns across devices. To address this trade-off, this study develops a trend-clustered forecasting framework that enables model sharing among transformers with similar temporal behaviors. Representation Learning for Time Series using Time-Embeddings (T-Rep) is employed to learn compact temporal representations of transformer load series, which are subsequently grouped using a Gaussian mixture model (GMM). Representative transformers are sampled within each cluster to train cluster-specific Patch Time Series Transformer (PatchTST) models while retaining device-level predictions. The framework is evaluated using real-world 15-min operational data from nearly one thousand distribution transformers. The proposed framework reduces mean absolute error (MAE) by approximately 20% relative to the direct PatchTST baseline, achieves a high-load warning accuracy of 95.66%, while requiring only seven cluster-specific forecasting models. These results demonstrate that behavior-based device organization provides an effective balance between forecasting accuracy, deployment scalability, and risk-oriented operational support for distribution transformer management.

arXiv:2611.11169v1cs.CV

Map2SLAM: From Streaming Map Prediction to Real-Time SLAM

Streaming 3D foundation models (3DFMs) estimate camera poses and dense geometry online, but they fall short of SLAM: their predictions are fixed once emitted, so loop closures cannot correct accumulated drift. With such models already providing strong geometric capability, a central question for SLAM becomes how to schedule their inference and how to keep its results as revisable state. We present Map2SLAM, which keeps LingBot-Map frozen, trains only a lightweight retrieval head, and builds SLAM on the model's inference interface, turning its predictions into observations of an external persistent SLAM state. The front-end tracks every incoming frame with the model's causal streaming inference and decides which frames are written to the persistent state as keyframes. The back-end asynchronously decides when and which past frames to re-observe, re-running the model's joint inference over them with fresh context; verified loop observations correct the trajectory online, and segment observations refine dense geometry after the sequence. All invocations share a single frame encoder pass per image, whose cached features also drive loop retrieval, so frequent re-invocation incurs no additional encoding cost. Map2SLAM achieves state-of-the-art performance on KITTI, VBR, and Oxford Spires, reducing mean ATE by over 40% compared with the best uncalibrated baselines, while maintaining real-time tracking at over 28 FPS at a resolution of $518\times154$ on an NVIDIA RTX 5090D GPU. Please visit our project page at https://robbyant-research.github.io/Map2SLAM/.

arXiv:2611.11165v1cs.LG

The Race Is Decided Early! A Kernel Law Predicts Lock-In in Majority-Vote Test-Time RL

Test-time reinforcement learning with majority-vote rewards trains a language model on unlabeled questions by rewarding the samples that agree with its most frequent answer. During training, each question locks into one answer, on hard questions often a wrong one, but which answer it locks into is known only after training. In this paper, we show that this outcome is decided early, in a short race driven by sampling noise between the answers that are close at the start, and we derive the probability of each outcome. With a separate logit for each answer, the runner-up wins with a normal tail probability of the initial gap, scaled by an explicit noise constant. In a network, however, the answer logits depend on shared parameters, and through the kernel of these parameters the updates of the trailing answers also change the gap of the two leaders. This spillover shifts the boundary between the starts won by each leader, and a closed form computes the shifted probability from a kernel probe and the start logits. In LoRA experiments with three models from two families, predictions made before training with the kernel frozen at its start value forecast which answer each question locks into better than a logistic model refitted on the outcomes, although they use no fitted constants. The closed form matches these predictions on the runs in which a check before training finds that the two leaders stay above the other answers, and it explains why a law that ignores the spillover needs a strongly inflated step.

arXiv:2611.11152v1cs.LG

Full Fine-Tuning is not the Problem: Why Adam Fails and SGD Succeeds in Few-Shot CLIP Adaptation

Few-shot CLIP adaptation is often treated as a parameterization problem. Full fine-tuning is assumed to overfit, so parameter-efficient fine-tuning (PEFT) methods are preferred. In this paper, we show that this conclusion conflates parameterization with optimization. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, dense full fine-tuning with plain SGD outperforms representative PEFT methods under the reported evaluation protocol, whereas Adam/AdamW collapses. We trace this collapse to unreliable support-set preconditioning. The few-shot second moment systematically underestimates the corresponding full-data statistic, already at initialization. Adam's inverse-square-root normalization converts this error into an oversized adaptive update. Matching Adam's first-step update budget to SGD resolves most of the collapse, whereas numerator-side alignment does not explain the failure. Shampoo and SOAP also fail, so richer preconditioning alone does not resolve the problem. Theory formalizes how few-shot estimation error is amplified into unsafe updates. SGD succeeds because it avoids unreliable preconditioners and stays close to pretrained CLIP. Interpolation probes show low-loss connections along tested directions, and curvature probes show favorable directional curvature. These results identify optimizer-induced preconditioning, not dense full fine-tuning itself, as the main source of failure in few-shot CLIP adaptation.

arXiv:2611.11151v1cs.DS

Fast Approximate $\ell_p$ Chamfer Distance via Lopsided Embeddings and Structured JL

For two $d$-dimensional point sets $A, B$ of size $n$, the Chamfer distance from $A$ to $B$ is defined as $\text{CH}(A,B)=\sum_{a \in A} \min_{b \in B} \mathsf{dist}(a, b)$, where $\mathsf{dist}$ is some underlying distance measure. We present efficient algorithms for approximating the Chamfer distance under the $\ell_p$ norm for $1 < p \le 2$. For the general case $1 < p < 2$, we utilize lopsided $\ell_p$ to $\ell_1$ embeddings with weak guarantees. We show that they are sufficient for preserving the Chamfer distance. In the high-dimensional regime, we use Fast Matrix Multiplication techniques to speed up the lopsided embeddings. These lead to $O(nd\cdot \mathrm{poly}(\log \log n, \log(1/\varepsilon))/\varepsilon^2)$ total time. For the specific Euclidean case ($p=2$), we leverage the Fast Johnson-Lindenstrauss Transform based on Toeplitz matrices and re-analyze the previous $\ell_1$-specific Chamfer algorithm in $\ell_2$. These achieve a runtime matching the recent state-of-the-art $\ell_1$ result of $O(nd(\log \log n + \log (1/\varepsilon))/\varepsilon^2)$, improving upon the previous $O(nd \log n/\varepsilon^2)$ bound for $\ell_2$. We also give additional results for the angular distance and upper and lower bounds in the streaming setting.

arXiv:2611.11149v1cs.LG

ZarrFM: A Canonical Tensor Storage Architecture for EEG Foundation-Model Pretraining

Statistical models for brain-computer interfaces (BCIs) increasingly require large, heterogeneous EEG corpora, yet existing datasets vary in sampling rates, channel layouts, window lengths, and storage formats, creating substantial preprocessing and I/O overhead. We propose ZarrFM, an execution-oriented representation that complements BIDS by storing a corpus as a single Zarr v3 hierarchy with indexed metadata, canonical channel spaces, explicit masks, hash-stable subject splits, verified event tables, and block-aligned sharding. On real PhysioNet recordings and synthetic corpora, we compare ZarrFM with EDF, FIF, NPY, HDF5, and Parquet. Across 200 recordings, its two-level index reduces metadata queries from 601 requests to 14. Preserving native signal codes reduces storage from 42.4 MB/hour in EDF to 22.7 MB/hour, while block-aligned sampling reaches 3,131 windows/s, up to 11 times faster than Parquet. Reads remain bit-identical across independent Zarr implementations, and concurrent failure-injected appends recover without data loss. These results show that ZarrFM can provide a compact, interoperable, and execution-efficient storage layer for large-scale EEG pretraining.

arXiv:2611.11147v1cs.CV

Not All Layers Are Equal in Image-to-Video Transfer

Parameter-efficient image-to-video transfer adapts frozen image foundation models to video via lightweight trainable adapters. Existing methods insert the same adapter at every transformer layer, implicitly assuming all layers require equal temporal modeling. A layer-wise sensitivity analysis across four video-native model families and two pretraining paradigms reveals that this assumption is wrong: temporal sensitivity increases with depth, and the pretraining paradigm determines its growth profile. Vision-only models exhibit a gradual rise, while vision-language models show early suppression followed by late recovery. This finding motivates STRAP, a training-free adapter placement rule that derives a spatial-to-temporal transition boundary from a paradigm-matched video model. Spatial adapters are assigned before this boundary and temporal adapters after it, without modifying the adapter architecture itself. On four fine-grained action recognition benchmarks, STRAP improves accuracy by up to 11.8% over uniform placement, reduces trainable parameters by 29 to 48%, and lowers inference time by up to 10%. On standard benchmarks, it matches accuracy with fewer parameters. Code is available at https://github.com/th-nesh/STRAP.

arXiv:2611.11145v1cs.LG

DISCOVER: Online Variance-Guided Data Discovery for Budgeted Multimodal GRPO

Recent advances in reinforcement learning have made verifiable rewards a central post-training paradigm for improving multimodal reasoning. However, methods such as GRPO remain data- and compute-inefficient, spending rollout and annotation budget on prompts that provide little learning signal. We propose DISCOVER, an online data-discovery framework for budgeted multimodal GRPO that enables vision-language models to dynamically identify prompts that are most useful for the current policy. DISCOVER first organizes the multimodal sample pool into semantic clusters and queries a small set of density-diverse representative samples to obtain initial reward observations. During training, it estimates the utility of unannotated samples by transferring sparse reward feedback from queried samples. Since sample utility changes as the policy evolves, DISCOVER dynamically updates these estimates with recency-weighted rollout observations. It then selects only high-utility prompts for annotation using a policy-conditioned score that accounts for reward variance and difficulty. Once the discovery budget is exhausted, DISCOVER continues with utility-based replay. Experiments on ViRL39K and VLAA-Thinking with Qwen3-VL-2B show that DISCOVER substantially improves multimodal mathematical reasoning under a strict 10% unique-sample budget, outperforming budget-matched random selection, hard-example sampling, full-pool GRPO under the same training-step budget, and online selection baselines. Ablations further validate the importance of semantic coverage, online reward transfer, and stable discovery cadence, establishing DISCOVER as a practical framework for data-efficient multimodal RL post-training.

arXiv:2611.11141v1cs.CE

TALK: One-Shot Batch Design for Protein Variant Effect Prediction

Protein variant effect prediction is central to understanding protein function and guiding protein engineering, yet experimental labels remain costly. In many protein experiments, the variants that can be assayed are dictated by protocol constraints, while the variants that need accurate prediction may lie in a different or broader region of the landscape. This assay-prediction mismatch makes candidate-local criteria such as predicted function, uncertainty, or exploration value insufficient for measurement design. Given a protocol-defined accessible space and a limited budget, we ask which variants should be assayed so that the resulting data best improves landscape prediction. We propose TALK, which estimates each candidate's contribution through posterior coupling between accessible and target variants and reduces redundant information within the batch. Across retrospective ProteinGym and FLIP protocols, TALK generally improves target-space prediction under limited measurement budgets. These results point to a practical two-stage workflow: first, spend a limited assay budget on variants selected to be informative for the target landscape and train a predictive model from the measured labels; then use this model to score and prioritize high-potential variants. Code is available at https://github.com/joehuang618/TALK.

arXiv:2611.11140v1cs.CV

SurgWorld: Learning Surgical Representations Beyond Observation via Latent Prediction

Surgical procedures are inherently partially observable, requiring surgeons to continuously infer evolving anatomical dynamics and future surgical states beyond the current observation. However, existing self-supervised learning methods primarily learn from observed surgical videos through masked reconstruction or predicting representations within the camera's current field of view, lacking the capability to explicitly model the out-of-sight surgical dynamics. Inspired by latent predictive world models, we introduce SurgWorld, a video-native, self-supervised framework for surgical representation learning, which treats a specified spatial or temporal shift as a pseudo-action and learns to predict the target latent representation from a partial observation. Specifically, we propose Out-of-Sight Dynamics Modeling (OSDM) as the pretraining objective for predicting representations conditioned on relative spatio-temporal shifts. We further regularize the learned latent dynamics from OSDM with a Geometric Visual Similarity (GVS) objective by aligning patch-level relational structures between predicted and target representations, preserving spatial geometry and local visual coherence for dense perception. These two objectives are integrated through an agglomerative training framework that consolidates complementary supervisory signals from multiple teachers to learn a unified latent representation of surgical dynamics and visual geometry. Extensive transfer learning experiments on 17 surgical benchmarks show that SurgWorld consistently outperforms existing vision foundation models across surgical workflow recognition, fine-grained action understanding, and dense visual perception. We further conduct tailored qualitative and quantitative analyses, demonstrating that SurgWorld effectively captures surgical dynamics beyond the directly observed context through latent prediction.

arXiv:2611.11139v1cs.CV

Graphene: A Lightweight Transformer with Ising Mean-Field Refinement for Image Super-Resolution

This paper proposes Graphene, a lightweight image super-resolution architecture with two complementary feature refinement mechanisms: Ising-model-inspired local feature refinement using an Ising Refinement Block (IRB) and stage-history-based feature refinement using a History Attention (HistAttn) module. The IRB reuses attention output features and pre-softmax attention logits to construct signed pairwise token couplings within each local window, applies unrolled mean-field updates, and refines features through a gated residual correction. These signed couplings allow alignment and opposition between latent spins to modulate local feature refinement. HistAttn attends over the current stage output and earlier representations at the same token position and adaptively blends the current stage output with a history-weighted feature aggregate without adding spatial pairwise attention. On Urban100 at $\times4$, Graphene and Graphene-light improve over IET and IET-light by 0.24 dB and 0.38 dB, respectively, with 34.3%/13.7% fewer parameters, 44.8%/44.4% fewer FLOPs, 78.4%/71.5% lower latency, and 76.0%/85.5% lower peak memory. Real-world SR comparisons further show that Graphene recovers fine details and preserves textures across diverse degradations. Code and reference checkpoints are available at https://github.com/skethz/Graphene