PyTorchKR Papers

최신 논문

총 568,830편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.11951v1cs.CV

UltraVoxelGS: Voxel-First Feed-Forward Gaussian Splatting for 3D Ultrasound Reconstruction

We present UltraVoxelGS, a 3D ultrasound reconstruction method that achieves both fast reconstruction and physically faithful rendering for the first time, bringing live spatial feedback during 2D ultrasound scanning within reach. This is made possible by a feed-forward design that predicts a complete Gaussian scene representation from posed ultrasound images directly, entirely bypassing the per-scene optimization bottleneck that has confined prior methods to offline use. Realising such feed-forward prediction in this setting is non-trivial: unlike natural images, ultrasound slices share at most one-dimensional intersections and exhibit view-dependent appearance due to acoustic propagation and attenuation, violating the dense overlap and photometric consistency assumed by existing feed-forward Gaussian methods. Hence, to achieve fast reconstruction, we introduce a feed-forward voxel-to-Gaussian pass that decouples representation size from input count by lifting all posed slices into a fixed-size voxel support and predicting Gaussian parameters directly. For faithful rendering, we propose ultrasound-aware appearance rendering strategy factorize reflection from cumulative attenuation during training and exploit appearance continuity among spatially adjacent slices for efficient Render-time Appearance Adaptation. On both real-world and simulated datasets, UltraVoxelGS achieves best PSNR/SSIM compared with per-scene optimization baselines while reducing inference from 10-20 minutes to approximately 10 seconds (over 35 times faster) and scales to sequences exceeding 1,000 frames. Code is available at https://github.com/YuebyYue/UltraVoxelGS-camera-ready.

arXiv:2611.11949v1cs.CV

Watching Itself Watch: Self-Auditing Visual Reliance for Video Reasoning

Video reasoning hinges on a small number of visually decisive tokens embedded in otherwise largely language-driven chains of thought. Yet prevailing post-training methods such as GRPO overlook this structure, assigning coarse trajectory-level rewards while requiring multiple rollouts and often costly external judges. We introduce Visually-Aware On-Policy Self-Distillation (VIOS), a post-training framework that identifies where a model under-utilizes visual evidence and concentrates learning at those positions. Our key insight is that a model can audit its own visual reliance: the divergence between its next-token distributions with and without the video reveals how strongly each prediction depends on the visual input. Measuring this reliance for both a student and an answer-privileged teacher exposes a per-token reliance gap, identifying positions where the teacher exploits visual evidence that the student overlooks. VIOS translates this signal into dense, targeted supervision through three complementary components: an exponential moving average (EMA) teacher that co-evolves with the student, a contrastive objective that encourages predictions to depend on the video, and a reliance-gap gate that focuses learning on visually under-utilized tokens. Crucially, VIOS requires only a single on-policy rollout from a single model, without external rewards, judges, or larger teachers. Across 12 benchmarks spanning general, temporal, spatial, knowledge-intensive, and long-video understanding, VIOS consistently improves three model backbones and outperforms contemporary video-reasoning systems, while training in over 9x fewer steps, 60x fewer FLOPs, and 111x fewer generated tokens. Project page: https://vision.cs.utexas.edu/projects/VIOS/

arXiv:2611.11947v1cs.RO

The Relative Value of Demonstration Data: Execution Time Limits and Operator Mixtures in Behavior Cloning

Robot demonstration datasets combine demonstrations from teleoperators of uneven experience, so collection teams need to know how many demonstrations from one group match those from another. We measure this with a data multiple: the ratio of the numbers of demonstrations two groups need for their policies to reach the same success rate. On robomimic's simulated Square and Can tasks, we train behavior-cloning policies on demonstrations from the two teleoperators robomimic labels Better and the two it labels Worse, varying dataset size and the share of Worse demonstrations. On Square, within robomimic's 25-second time limit, an all-Worse training set needs 1.55 times as many demonstrations as an all-Better set. Most of this difference is late completion: policies trained on Worse demonstrations finish later, as the Worse teleoperators do, and with a 100-second limit the multiple falls to 1.11. On Can, most of the difference remains at 100 seconds. If each Worse demonstration were worth 1/1.55 of a Better one, as in an all-Worse set, a half-Better, half-Worse set on Square would need 1.22 times as many demonstrations as an all-Better set. In fact, it needs 13% more than that, and this excess persists at every time limit from 20 to 200 seconds. The excess remains when training draws demonstrations uniformly instead of in proportion to their length. On Square, within the 25-second limit, switching to these Worse teleoperators is cheaper at equal success only if they earn less than a third of the Better teleoperators' hourly pay, because Worse demonstrations also take longer to record.

arXiv:2611.11946v1cs.LG

Decision Aware Origin-Destination Demand Estimation with Structurally Constrained Data

Origin-destination (OD) demand plays an important role in transportation system design and management applications. Conventional OD estimation methods typically judge estimation quality based on how estimated demand aligns with observed traffic data and historical demand patterns. We argue that data alignment alone is insufficient; how errors in OD estimates influence downstream decision quality should also be considered. This paper develops a decision aware OD demand estimation framework that leverages structurally constrained traffic data to characterize an admissible demand region, within which we identify OD estimates that would lead to least regret in the downstream decision. Therefore, our OD estimates are both data-consistent and decision-relevant. Methodologically, we approximate the admissible demand region induced by physical network constraints through a geometric representation, which regularizes the search over feasible OD estimates. Within this regularized region, we use a convex surrogate of downstream regret to obtain a tractable decision aware optimization problem. Numerical experiments show that the proposed framework consistently reduces decision regret relative to conventional estimation benchmarks, demonstrating the value of aligning OD demand estimation with downstream decision quality.

arXiv:2611.11937v1cs.LG

Attention Is Only as Good as Its Witnesses: When Higher-Order Layers Help, and When They Do Not

Transformers stop learning graph search as graphs grow, and learn the card game SET only after many distinct questions, although attention can represent both. We read these failures, and the uneven record of higher-order attention, from what a layer sums over: each layer we study computes the partition function of a family of witnesses and reads it at a temperature. Two laws follow. The depth law is a known result of logic, which we state for these layers: L layers over two-step walks reach exactly 2^L - 1 steps, just as a logical statement with L nested clauses of the form "there is a vertex such that" sees exactly that far. We build a construction that reaches this far, checked in Lean, and, when vertices carry no names, no such network reaches further. The size law is the cost of reading at a temperature, which logic does not have: existence is read correctly at every size, in the worst case, if and only if inverse temperature times margin is at least the logarithm of the number of candidates. Layers built on these laws, and on two principles that we test, learn where Transformers fail. One propagator layer trained on 49-vertex graphs answers every question at every search depth on 250-vertex graphs. Given the modulus 3, exponential Newton computes the softmax of 2-simplicial attention over pairs exactly in 3^d n operations for d clocks; with six clocks it learns SET from as few deals as 2-simplicial attention and transfers from 8- to 20-card tables. On CLRS-30 without hints, a 186,000-parameter block of OR-join and power-mean channels matches (within half a point) or beats a model of at least 73 million parameters on 9 of 12 graph tasks of the official test set, and on 8 of 12 with freshly drawn test graphs. The same account explains where pair layers did not help and where parallel layers must loop.

arXiv:2611.11936v1cs.LG

Constrained Variational Policy Optimization for Hybrid Action Spaces

Constrained reinforcement learning in hybrid action spaces must coordinate discrete modes and mode-conditioned continuous controls under long-horizon cost constraints. Existing constrained hybrid-action methods typically optimize factorized policies directly in parameter space, whereas Constrained Variational Policy Optimization (CVPO) separates nonparametric improvement from parametric projection in its original continuous-action formulation. We introduce Hybrid Constrained Variational Policy Optimization (H-CVPO), extending this variational framework to hybrid action spaces. Under boundedness and strict feasibility, its constrained E-step has zero duality gap and, when the dual optimum is attained at positive temperature, yields a closed-form joint target distribution over hybrid actions through exponential reward-cost reweighting. Using the exact hybrid KL decomposition, we formulate a factorized M-step that fits both policy components to this target under separate discrete and conditional-continuous trust regions. For exact E/M updates, we prove fixed-reference surrogate improvement and conditional reward-cost bounds with a sufficient feasibility-margin condition. For approximate updates, we establish the existence of a dual minimizer under continuity on a compact domain and use residual-based admissibility to determine whether to project the target. Ablations identify cost-aware reweighting as an important contributor to constraint control. Extensive experiments in simulated hybrid-control environments show that H-CVPO achieves favorable reward-cost trade-offs compared with hybrid-action baselines.

arXiv:2611.11935v1quant-ph

AI-Driven Discovery of Entanglement Witness in Spin Hamiltonians

Entanglement is a crucial resource in quantum science and engineering, but detecting it in many-body states from experimentally accessible observables remains challenging. Here we introduce EntNet, an AI-driven framework to discover compact combinations of spin observables that accurately distinguish entangled from fully separable many-body mixed states in one-dimensional spin-$1/2$ chains governed by short-range Hamiltonians. We construct local-unitary-invariant representations of measurable Pauli correlators up to four-body order, sparsely select 16 candidate invariant feature channels, and learn their spatial structure using a Transformer architecture with rotary positional embeddings. Using only expectation values of these low-order observables, EntNet achieves $>98\%$ recall on certified entangled states with near-zero false-positive rates on separable states. Remarkably, it exhibits transferability across system sizes and distinct Hamiltonian families, including frustrated and disordered spin models, with minimal loss of accuracy. Beyond detection, EntNet provides physical insight into entanglement detection by revealing a selective spatial organization of pair correlations, while its finite-temperature decision boundary tracks the excitation gap of Jordan-Wigner quasiparticles. These results establish a transferable and physically interpretable framework for data-driven discovery of entanglement witnesses from local correlation functions.

arXiv:2611.11934v1cs.CV

Scale-Consistent Monocular Fisheye Visual Odometry via Spherical Bundle Adjustment-Guided Unsupervised Depth Learning

The wide field of view of fisheye cameras and the simplicity of a monocular setup make monocular fisheye visual odometry (VO) attractive for robotic localization and mapping. Nevertheless, scale drift remains a major challenge over long trajectories. Unsupervised depth learning can help to mitigate this issue, but the pronounced distortion of fisheye images poses challenges for accurate depth estimation. This article presents S$^2$D-FVO, a monocular fisheye VO framework integrating spherical bundle adjustment (SBA)-guided unsupervised depth learning to improve scale consistency and depth estimation accuracy. Given that pixel reprojection residuals do not uniformly reflect angular deviations between viewing rays and that the first-order linearization of a scalar angular reprojection residual is insensitive to perturbations along one tangent direction, SBA employs tangent-space angular reprojection residuals derived from the spherical logarithmic map. These residuals retain sensitivity along both tangent directions, with their norms equal to the corresponding scalar angular residuals. A median depth-based scale regularization term is further integrated into SBA to constrain the reconstruction scale. Building on these formulations, SBA replaces the pose estimation network during unsupervised depth learning, providing more reliable relative poses for view synthesis. In turn, depth estimates from the well-trained depth estimation network provide scale references for SBA, improving scale consistency while allowing individual 3-D landmark positions to be refined. Extensive experiments conducted on multiple public datasets, as well as two newly collected large-scale datasets, demonstrate that the proposed framework outperforms representative existing approaches in both depth and trajectory estimation accuracy and is applicable to different camera models.

arXiv:2611.11929v1cs.AI

Can LLM Agents Organize Their Own Memory?

In most LLM agents, weights are fixed after deployment, so new information must be retained in external memory. Existing approaches often rely on predefined records and fixed organizational rules, which cannot readily accommodate evolving tasks and interactions. This raises the question: can LLM agents organize their own memory using ordinary documents? We introduce self-organized memory (SoMem), a framework that stores memory in linked documents. The agent decides what to summarize, how to group related information, and when to revise existing topics, and every summary it produces stays linked to the original interactions as evidence. The runtime validates updates and automatically derives temporal, relational, and recent-memory views from the organized content. At query time, the agent chooses what to search and read, following document links and inspecting original evidence as needed. On LoCoMo, BEAM-100K, and LongMemEval-S, SoMem outperforms the strongest baseline by 2.66, 3.10, and 2.00 percentage points, respectively. The constructed memory can be reused across models, and additional reasoning effort can further improve answer quality. These findings demonstrate the feasibility of agents organizing and using their own persistent memory through ordinary documents. Code is available at: https://github.com/Qi202/SoMem.

arXiv:2611.11927v1q-bio.NC

Language processing in the autistic brain: A meta-analysis of neuroimaging studies across modalities and age

Language is central to social communication, which is altered in autism, and language abilities vary widely among autistic individuals. How the autistic brain processes language is central to understanding autistic communication. Yet neuroimaging studies report inconsistent findings on how autistic brain activity is altered during language processing. We synthesized 45 fMRI and PET studies (753 autistic and 777 non-autistic participants) with seed-based d mapping with permutation of subject images (SDM-PSI). We estimated autistic and non-autistic activation and their difference with meta-regressions adjusted for each study's language modality and mean age. Autistic and non-autistic brains activated a similar left-lateralized language network. Autistic brains showed reduced activity throughout the left perisylvian cortex, from the inferior frontal gyrus to the mid superior temporal gyrus and the posterior superior temporal sulcus, and increased activity in the right temporal pole. Spoken language showed a larger autistic reduction than written language in the left mid superior temporal gyrus. Adults, but not children and adolescents, showed increased autistic activity in the right superior temporal gyrus and cuneus. Autistic language processing thus relies on the typical left-lateralized network but shows reduced leftward lateralization.

arXiv:2611.11926v1cs.CV

Text Alone Predicts Which Classes a Segmentation Readout Keeps

An open-vocabulary segmentation readout silently loses a class deployed beyond its training categories. It still emits masks with high confidence, but the class never wins a single pixel. The training loss falls anyway, and mIoU can even rise while classes die. We show that which classes collapse is written in the class-name list alone. Before any image is seen, we rank each class's collapse risk from its text similarity to the training vocabulary alone. The ranking holds at AUROC 0.69-0.96 across nine model families, twelve backbones and eight pretraining recipes. Two further families, measured only after the predictions were written down, land in the same band. What the list writes, it can rewrite. With identical images and a class's counterpart moved in or out of the training list, retraining flips that class's gain over the frozen baseline by 21 IoU points; swapping in unrelated names changes nothing. We find a vocabulary-distance mechanism behind the collapse. Collapsed classes lose every vote to near-name neighbors, but in the frozen features they survive. Turning down the readout's amplitude, or stopping training early, brings them back, up to +16 points, without retraining; trained with its amplitude held down, the readout never collapses.

arXiv:2611.11923v1stat.ME

Characterizing Underrepresentation when Generalizing Causal Survival Estimates

Randomized trial findings are routinely generalized to broader target populations to inform decisions in medicine and public policy. When the trial sample is misaligned with the target population, generalized effect estimates can yield suboptimal decisions. We characterize the target subgroups that are underrepresented in the trial cohort in time-to-event settings. Here, underrepresentation arises in two distinct ways: baseline underrepresentation, when a subgroup's covariate profile is rare in the trial relative to the target, and time-varying underrepresentation, when differential censoring erodes a subgroup over follow-up. Existing diagnostics either conflate these mechanisms or apply only to the baseline case. We show that the asymptotic variance of the efficient generalizability estimator admits a decomposition in which only two terms can diverge as representation weakens: a sampling-weight term capturing baseline misalignment and a censoring-weight term capturing differential loss to follow-up. Underrepresented subgroups are therefore those with imprecise generalized estimates, and the larger of the two terms gives the responsible channel. We use this decomposition in TRACE (Temporal Rashomon Analysis of CEnsored data), which combines a self-normalized variance objective with a tree-based search to localize underrepresented subgroups, attribute each to baseline mismatch or censoring, and find the follow-up intervals where censoring drives the variance inflation. We solve the joint problem with a parametric dynamic program nested inside a Rashomon ensemble of near-optimal trees. In synthetic benchmarks, TRACE recovers the planted underrepresented regions. In a case study generalizing the National Job Training Partnership Act (JTPA) Study to U.S. JTPA-eligible adults, restricting to reliably generalizable subgroups narrows the 95% confidence interval for the refined estimand by 60%.

arXiv:2611.11922v1gr-qc

Deep-Learning-Assisted Bayesian Inference for Extreme-Mass-Ratio Inspirals

Extreme-mass-ratio inspirals (EMRIs) pose a major challenge to Bayesian parameter inference because of their complex, highly multimodal likelihood surfaces. Locating narrow high-likelihood regions is particularly difficult without knowledge of the true source parameters, which is unavailable in real observations. We develop a deep-learning-assisted Bayesian inference framework that combines neural-network-guided parameter-space localization, particle swarm optimization, and Markov chain Monte Carlo (MCMC) sampling. Applied to simulated six-month EMRI signals, the framework recovers posterior distributions in close agreement with reference MCMC analyses initialized near the true parameter values, without requiring such knowledge for initialization. Once the neural network has been trained, the complete inference pipeline takes approximately 10 hours per signal.

arXiv:2611.11921v1cs.LG

An Evaluation of Tabular Foundation Models as Training-Free World Models

World models are widely considered to be a potential key component of model-based reinforcement learning and robot planning, but typically require task-specific training on substantial interaction data. We investigate whether a pretrained tabular foundation model (TabFM) can serve as a training-free dynamics model for world modeling by framing next-state prediction as in-context regression over transition pairs, and where such a model helps and where it breaks. Across synthetic dynamical systems and robot locomotion benchmarks, TabFM is more accurate than LSTM and MLP models trained from scratch on the same limited transitions in almost all settings, and among regressors fitted to the same data it is the most accurate on locomotion, while a Gaussian process is more accurate on smooth low-dimensional systems. However, models trained on the full dataset remain substantially better on most systems. When rolled out autoregressively over multiple steps, TabFM's accuracy depends strongly on how much context it is given. With a short context, its rollouts quickly degrade to the level of a trivial baseline, whereas with a long context they improve substantially and, on one locomotion task, exceed a trained LSTM. Inference is also about three orders of magnitude slower than an LSTM. Overall, in their current state, we find that training-free priors are useful as predictors when data are scarce and enough context is available, but they are not yet a substitute for trained world models in planning.

arXiv:2611.11916v1cs.LG

LBSNet: Lightweight Brain State Modeling via Learnable Spectral Prototypes for EEG-Based Fatigue Detection

EEG-based fatigue detection requires modeling spectral information across bands and its temporal variation. Recent methods combine multi-band modeling with adaptive temporal processing, yet how spectral structure across bands should influence the contribution of each temporal interval remains insufficiently explored. We propose LBSNet, a lightweight network that uses prototype-based brain-state modeling to link local spectral structure with temporal aggregation. LBSNet extracts compact multi-band features through band-specific temporal convolution, shared spatial projections, and cross-band recalibration. Within each window, relative spectral activity is encoded as a normalized descriptor and softly matched to learnable spectral prototypes. Prototype assignments and learned prototype importance jointly determine temporal weights for power aggregation, allowing the spectral pattern of each window to guide its contribution to the prediction. Subject-independent experiments on SEED-VIG and SADT show that LBSNet outperforms seven baselines while maintaining low computational cost. Ablation and visualization analyses support prototype-guided temporal weighting, while the learned prototypes reveal a gradual redistribution of relative cross-band spectral states with increasing fatigue. The code is publicly available at https://github.com/zy0929/LBSNet.

arXiv:2611.11914v1cs.LG

CRAFT-GBM: Selecting Best Features for Gradient-Boosted Machine Learning

Feature subsampling is used to accelerate gradient-boosted trees, yet standard approaches fix the feature count in advance and largely ignore the split-gain evidence accumulated during training. We introduce CRAFT-GBM (Confidence-Ranked Adaptive Feature Tracking), a native LightGBM extension that uses this evidence to adapt which---and, by default, how many---features are searched each boosting round. CRAFT-GBM transforms root-split gains into bounded observations and constructs joint time-uniform bounds valid under adaptive, non-stationary training histories. In its default endogenous mode, after initialization, CRAFT-GBM retains exactly the features that cannot yet be ruled out as best, letting the selected feature-set size $K_t$ emerge from training history rather than a pre-specified fraction. With probability at least $1-\delta$, every feature with the largest expected gain is retained jointly across all rounds after initialization. When a feature fraction $f$ is specified instead, CRAFT-GBM ranks features by their interval point estimates and provides an approximate top-$K$ guarantee for feature budget $K_f$. We implement CRAFT-GBM within LightGBM while preserving its native feature-bundling representation, and evaluate it across 14 public and synthetic dataset families and four noise-augmented settings. In the fully endogenous setting, CRAFT-GBM reduces per-fit training time by 39.74\% with 1.04\% lower test loss in the geometric mean across all 18 RQ1 experiments. Under pre-specified feature fraction $f$, Exogenous CRAFT-GBM also yields aggregate training-time reductions up to 13.52 percentage points larger than random feature subsampling, with no increase in aggregate test loss relative to full-feature LightGBM. Accumulated training evidence can thus support adaptive feature search with statistical guarantees while retaining the practical advantages of native gradient boosting.

arXiv:2611.11913v1cs.AI

CHORD: Multi-resolution Compression for Multimodal Time Series

Wearables, mobile devices, and IoT platforms generate multimodal time-series data continuously and at scale, making efficient storage and transfer a primary concern. The same recordings are reused after collection by many downstream models, often at different temporal resolutions, so a compressed representation should be small, useful across resolutions, and produced at a bit rate known in advance. Prior work meets parts of this: symbolic and error-bounded compressors reconstruct accurately but fix resolution at compression time, learned codecs target one signal domain, and discrete tokenizers reconstruct without accounting for rate. We introduce CHORD, a coarse-to-fine compressor for heterogeneous multimodal time series. A residual cascade orders each payload from broad to fine temporal scales and synthesizes corrections through learned shapelet-inspired atoms. A shared pathway supplies a reusable synthesis basis across modalities, while private pathways retain capacity for sensor-specific morphology. Finite scalar quantization fixes the bit width of every modality and scale, so payload size is known before encoding, and scale dropout trains each coarse-to-fine prefix to be a valid reconstruction operating point. Across seven wearable-sensing and forecasting benchmarks, CHORD stays within 0.62 points of uncompressed classification accuracy on average at the 8x target, the smallest gap of any compressed method, with all resolutions decoded from one payload.

arXiv:2611.11912v1cs.CV

LO-World: Unified BEV Latent Modeling for Joint 4D Occupancy and LiDAR Forecasting

LiDAR point clouds and 3D semantic occupancy represent two distinct modalities capturing fine-grained metric geometry and dense volumetric semantics, respectively, holding natural potential for mutual complementarity. However, existing world models typically forecast these modalities independently or cascade them heuristically, missing the opportunity for reciprocal cross-modal regularization. To address this, we propose \textbf{LO-World}, a unified 4D world model that forecasts both modalities within a shared BEV latent space. This single-stream formulation enables dense semantics and metric geometry to mutually reinforce each other, while eliminating multi-pipeline redundancy. Specifically, LO-World couples an efficient hybrid range-voxel autoencoder with a flow-matching generator that caches historical context to reduce iterative sampling latency. Extensive experiments demonstrate that LO-World outperforms prior state-of-the-art baselines, improving occupancy forecasting IoU by 10.9% and reducing LiDAR Chamfer Distance by 30.1%. Meanwhile, it sustains a real-time throughput exceeding $25$~FPS on a single RTX 4090. Our code is available at https://github.com/lzzzzzm/LO-World.

arXiv:2611.11910v1cs.LG

Late but Valuable: Why Deadlines Change the Measured Value of LLM Reasoning

Large language models (LLMs) increasingly plan and run experiments in automated laboratories. These experiments keep running while the model reasons and run on schedules with hard deadlines. If an LLM-generated answer arrives late, the system executes a safe fallback action instead, such as rejecting a sample. This creates a systems problem that standard serving metrics do not capture. An answer that matches the fallback adds no value, and answers that differ from the fallback may require more reasoning time. Reporting answer quality and latency separately can therefore overstate the value of reasoning. We test this on synthetic sample decisions, where the best choice is known exactly and a runtime enforces the deadline. We compare one LLM's fast mode, which answers directly, with its thinking mode, which reasons first. In a prespecified replay of 1,200 new decisions, thinking answers that missed a deadline were worth about five times as much as on-time ones. Multiplying the on-time rate by the mean answer value predicted about twice the value that thinking delivered. This prediction overstated thinking's advantage over fast mode by 0.0241 per decision, and by 0.0191 in a repeat on 300 new decisions. Both 95% confidence intervals excluded zero. In a single-GPU Qwen3-8B experiment using simultaneous request bursts, raising the number of requests from 8 to 56 reduced the share of thinking answers that met a 45 s deadline from 79.2% to 1.4%, although total output tripled. Evaluations of LLM agents for laboratories should therefore measure each answer's value and arrival time together and score the action that actually runs.

arXiv:2611.11907v1cs.AI

Layer Roles Matter: Mitigating Modality Imbalance in Multimodal Emotion Reasoning via Adaptive Layer Profiling

Multimodal Emotion Reasoning (MER) requires perceiving and integrating emotion cues from multiple modalities. Recent advances increasingly leverage Multimodal Large Language Models (MLLMs) to enhance reasoning capacity. However, due to the inherent heterogeneity in conveying emotions across modalities, these models often suffer from modality imbalance and confront flawed emotion understanding. In this paper, we conduct an in-depth analysis on MER models and uncover two key empirical insights. (i) In MER, modality imbalance manifests two distinct failure modes: unimodal over-reliance and insufficient cross-modal utilization, which call for different intervention strategies. (ii) MLLM layers exhibit implicit functional hierarchies in multimodal emotion processing, playing separable roles in stabilizing unimodal representations and integrating cross-modal information. Motivated by these observations, we propose a Layer-Aware Modality Rebalancing (LAMR) framework by explicitly aligning intervention strategies with layer roles. Specifically, we first introduce two quantized metrics for adaptive layer role profiling. In each layer, Modality Entropy (ME) measures the dispersion of modality contributions, and Cross-modal Intensity (CMI) captures the strength of cross-modal interactions. Then, we respectively enhance modality equilibrium in stabilization layers and cross-modal propagation in integration layers. Extensive experiments on two comprehensive MER benchmarks demonstrate that LAMR effectively mitigates modality imbalance and improves emotion reasoning across diverse MLLM backbones, without additional training.