Regularization
-
Enhancing In-context Panoramic Generation via Geometric-aware Pretraining
2026-07-09 17:59:06Haoran Feng, Ruiyang Zhang, Longyi Zhang, Dizhe Zhang, Lu QiAbstract
In this work, we present Canvas360, a two-stage framework for in-context panoramic generation that combines geometry-aware pretraining with downstream task-specific fine-tuning. To address the lack of large-scale, high-quality training data tailored to in-context panoramic tasks, we propose Canvas360Dataset, a collection of 1M high-quality paired panoramic samples for style transfer, inpainting, outpainting, and editing, enabling effective supervision across diverse in-context generation scenarios. On the modeling side, Canvas360 enhances text-to-panorama generation through parallel depth generation, velocity circular padding, and similarity loss regularization, enabling the model to learn geometry-aware representations, capture object distortion details, and improve geometric consistency and global coherence. Furthermore, empowered by strong panoramic priors, Canvas360 enables a unified in-context panoramic generation framework that supports diverse downstream tasks via token-level concatenation, surpassing prior methods in both task coverage and modeling flexibility. Extensive experiments show that Canvas360 improves panoramic image fidelity, achieving particularly strong performance on the panorama-specific FAED metric and competitive or leading results across the reported quantitative evaluations. More information can be found on our project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2607.08765
PDF
-
SLORR: Simple and Efficient In-Training Low-Rank Regularization
Abstract
Low-rank factorization is widely used to compress neural networks, but modern models are often not naturally amenable to aggressive factorization without significant accuracy loss. Existing training-time low-rank regularizers can improve compressibility, but they often require SVDs of large weight matrices, modify the model architecture (introducing additional trainable parameters), or rely on stateful cached quantities. To address these limitations, we introduce SLORR, a simple, stateless, and architecture-preserving framework for in-training low-rank regularization, instantiated with two main variants based on the Hoyer sparsity metric and the nuclear norm. SLORR directly regularizes the original weight matrices using GPU-friendly approximations for the forward and backward passes of the regularizers, for which we provide approximation guarantees. We first evaluate SLORR on ImageNet-1K across short-horizon continued training of ResNet-50, ViT-B/16, and ViT-L/16, and pretraining of ResNet-18, where SLORR induces compressibility while introducing less than 8% training overhead. We further evaluate SLORR-Hoyer in LLM pretraining at 135M and 560M scales: SLORR-trained compressed models preserve performance substantially better than unregularized models while adding less than 1% average training overhead.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08754
PDF
-
XFACTORS: Disentangled Information Bottleneck via Contrastive Supervision
2026-07-09 17:00:50Alexandre Myara, Nicolas Bourriez, Thomas Boyer, Thomas Lemercier, Ihab Bendidi, Auguste GenovesioAbstract
Disentangled representation learning aims to map independent factors of variation to independent representation components. On one hand, purely unsupervised approaches have proven successful on fully disentangled synthetic data, but fail to recover semantic factors from real data without strong inductive biases. On the other hand, supervised approaches are unstable and hard to scale to large attribute sets because they rely on adversarial objectives or auxiliary classifiers. We introduce \textsc{XFactors}, a weakly-supervised VAE framework that disentangles and provides explicit control over a chosen set of factors. Building on the Disentangled Information Bottleneck perspective, we decompose the representation into a residual subspace $\mathcal{S}$ and factor-specific subspaces $\mathcal{T}_1,\ldots,\mathcal{T}_K$ and a residual subspace $\mathcal{S}$. Each target factor is encoded in its assigned $\mathcal{T}_i$ through contrastive supervision: an InfoNCE loss pulls together latents sharing the same factor value and pushes apart mismatched pairs. In parallel, KL regularization imposes a Gaussian structure on both $\mathcal{S}$ and the aggregated factor subspaces, organizing the geometry without additional supervision for non-targeted factors and avoiding adversarial training and classifiers. Across multiple datasets, with constant hyperparameters, \textsc{XFactors} achieves state-of-the-art disentanglement scores and yields consistent qualitative factor alignment in the corresponding subspaces, enabling controlled factor swapping via latent replacement. We further demonstrate that our method scales correctly with increasing latent capacity and evaluate it on the real-world dataset CelebA. Our code is available at \href{this https URL}{this http URL}.
Abstract (translated)
URL
https://arxiv.org/abs/2601.21688
PDF
-
When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities
2026-07-09 15:35:08Weiduo Liao, Yunqiao Yang, Ying WeiAbstract
Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept. However, in vision-language models (VLMs), vanilla SAEs struggle to learn modality-consistent concepts, with concepts often exhibiting fragmented coverage (i.e., disjoint regions) in the visual modality. To address this challenge, we propose a Structured Sparse AutoEncoder ($S^2AE$) that enforces concept consistency from both semantic and spatial perspectives in the visual modality. Specifically, we group image patches based on Transformer attention similarity and spatial proximity, and introduce a structured sparsity regularization when training the vanilla SAE. The regularization consists of exclusive sparsity for inter-group concept disentanglement and group sparsity for intra-group concept consistency, which drives the latent neurons by SAEs to specialize in distinct, semantically grounded concepts. Evaluated on the \texttt{Qwen2.5-VL-7B-Instruct} model, the method achieves 6.06% average improvement in semantic alignment (mIoU) and 60.81 in representational efficiency (lower l0 norm) while maintaining near-perfect reconstruction fidelity with an Explained Variance above 99%. Cross-modal analysis further demonstrates that $S^2AE$ enhances neuronal monosemanticity by this visual structural prior, achieving a 3.08% average gain in semantic consistency and a 2.37% average gain in monosemanticity scores for both modalities of multimodal features, thereby fostering more coherent and disentangled representations.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08605
PDF
-
Ensemble Diversity Optimization for Subjective Supervision
2026-07-09 13:53:02Xia Cui, Ziyi Huang, N. R. AbeynayakeAbstract
Subjective NLP tasks often exhibit systematic annotator disagreement, requiring models that represent uncertainty rather than collapse it. We introduce Ensemble Diversity Optimization (EDO), a prediction-space framework that jointly optimizes ensemble weights, effective cardinality, and calibration through a unified differentiable objective. EDO learns ensemble composition and size end-to-end via Gumbel-Softmax relaxation and incorporates a signed diversity regularizer, tuned on validation data, to steer optimization toward either preserving or suppressing disagreement. This regularization prevents ensemble collapse and enables controlled navigation of the utility-calibration trade-off. The framework integrates a soft F1 surrogate, class-weighted cross-entropy to address imbalance, and reliability-weighted diversity to regulate intra-ensemble variability. Experiments on four subjective text-classification benchmarks (ArMIS, ConvAbuse, HS-Brexit, MD-Agreement) show that EDO substantially improves probabilistic calibration, reducing cross-entropy (40-78% depending on baseline) and lowering Brier scores relative to Soft-CE, Soft-MD, Top-5 Voting, and WEL, while maintaining competitive F1 and better alignment with annotator distributions. These results demonstrate that jointly optimizing ensemble structure with a signed diversity regularizer provides an efficient, model-agnostic approach for modeling human subjectivity in supervised learning.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08493
PDF
-
Feedback Manipulation Regularization: Enabling Offline Agent Alignment for Imitation Learning
2026-07-08 18:47:29Benjamin Poole, Minwoo LeeAbstract
Reinforcement learning (RL) research has increasingly shifted focus towards alignment, ensuring agents learn behaviors adhering to human values. While human demonstrations and feedback have proven crucial for alignment, existing approaches predominantly combine these signals using multi-stage pipelines designed for the contextual bandit framing of language generation. Yet little work explores how these complementary inputs can serve as a richer, interconnected signal for single-stage offline training in fully sequential decision-making environments. We propose Feedback Manipulation Regularization (FMR), an algorithm-agnostic method that harnesses evaluative feedback as a corrective signal to improve the alignment of imitation learning policies. We adapt Safety Gymnasium environments to be a principled testbed for alignment evaluation, demonstrating improved aptitude and up to a 98\% reduction in misalignment across a range of imitation learning algorithms. FMR remains robust in limited data regimes, even when learning from scarce aligned and uninformative noisy demonstrations.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07859
PDF
-
Graph-Regularized Deep Learning for EEG-Based Emotion Recognition with Psychologically-Grounded Label Structure
2026-07-08 16:41:51Dongyang Kuang, Zizheng Ma, Yushan Zhang, Xiaocong ZengAbstract
EEG-based emotion recognition is critical for mental health monitoring and affective brain-computer interfaces, yet existing deep learning approaches often treat emotion classes as isolated labels, ignoring their psychological interdependencies. We propose a graph-regularized learning framework that conceptualizes emotions as nodes in a graph where edges encode proximity based on dimensional emotion theories. We adapt three complementary regularization strategies--Graph Label Smoothing (intuitive soft labeling), Commuting distance on graph via Graph Laplacian (spectral graph theory), and Sliced Wasserstein Distance (optimal transport on graph)--ordered by increasing computational complexity. These strategies penalize model predictions that deviate from the established emotion topology. Our framework is evaluated across three representative backbone architectures: AudioTransformer (pure transformer), Conformer (CNN-transformer hybrid), and DCGNN (causal graph neural network), demonstrating architecture-agnostic benefits. Experiments on SEED-IV (4 classes) and SEED-V (5 classes) datasets show consistent improvements: best case up to +5.42% accuracy and 39% reduction in psychologically implausible misclassifications. Ultimately, our framework help raise the upper bound of performance achievable with standard approaches. Code will be released.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07773
PDF
-
VCDP: Variation-Conditioned Distributional Proxy Learning for Semi-Supervised Medical Image Segmentation
2026-07-08 13:46:53Zimu Zhang, Yiheng Zhong, Zhuoru Zhang, Yingzhen Hu, Yanan He, Fanliang Meng, Xiaofeng LiuAbstract
Semi-supervised 3D medical image segmentation reduces the need for dense voxel-level annotations by exploiting unlabeled volumes. Although existing methods such as consistency regularization, pseudo-labeling, and co-training improve prediction-level robustness, they often provide insufficient feature-space organization for anatomically complex structures, especially small organs and ambiguous boundary regions with large intra-class variations. To address this issue, we propose Variation-Conditioned Distributional Proxy Learning (VCDP), a plug-and-play training-only regularization module for semi-supervised 3D medical image segmentation. VCDP represents each class with a learnable Gaussian distribution for shared class semantics and multiple variation prototypes for fine-grained intra-class patterns. A unified variation-conditioned compatibility score is further formulated to fuse distributional similarity and soft variation aggregation, guiding voxel embeddings to align with both global organ identity and local anatomical variations. VCDP is attached to decoder features during training and removed during inference, introducing no additional inference cost. Experiments on multi-organ segmentation benchmarks show that VCDP improves most evaluated baselines, particularly for small, ambiguous, and highly variable organs. Our anonymous code is released at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07416
PDF
-
When Prompts Ignore Structure: Graph-Based Attribute Reasoning for Calibrated VLMs
2026-07-08 13:31:17Tanay Sodha, Aditya Sharma, Ramya Hebbalaguppe, Vinti Agarwal, Pranav Murthy YeluripatyAbstract
Reliable confidence estimation remains a key limitation of test-time adaptation in vision-language models (VLMs), where prompt tuning improves zero-shot accuracy but often degrades calibration due to entropy-driven overconfidence. Prior approaches mitigate this using LLM-derived class attributes and contrastive regularization, yet treat attributes independently, ignoring their relational structure. We propose ARGTCA, which represents (class, attribute) pairs as nodes in a Symbolic Attribute Graph and trains a Graph Attention Network (GAT) using contrastive objectives to produce structurally informed embeddings that capture inter-attribute dependencies. We introduce two attribute selection strategies: ARGTCA-DIV for intra-class diversity and ARGTCA-DISC for inter-class discrimination. Experiments across nine benchmarks show that ARGTCA-DIV reduces average Expected Calibration Error (ECE) by approximately ~37% over baselines, while ARGTCA-DISC consistently performs as the second-best variant, reducing average ECE by approximately ~17% over baselines. These results suggest that modeling symbolic attribute interactions provides a principled approach for reliable test-time adaptation in VLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07395
PDF
-
TACoS: Weakly Supervised Learning of Two-Dimensional Materials from Scribble Annotations to Precise Segmentation
2026-07-08 09:02:31Jiabei Chen, Liping Zhang, Jiang-Bin Wu, Zhongming Wei, Enhao Ning, Su Yan, Weijun Li, Ping-Heng Tan, Xin NingAbstract
The precise pixel-level localization of 2D material flakes is crucial for high-throughput screening. However, traditional fully supervised methods rely on dense annotations, which are costly and time-consuming, severely limiting the practical deployment of segmentation models. This paper proposes TACoS, a specialized scribble segmentation framework tailored for 2D materials. First, we design a unified framework that integrates semi-supervised consistency learning with structured tree energy constraints. This framework comprises two core components: an unlabeled weak-strong distribution alignment module and a tree energy regularization module. The former employs cosine consistency constraints to enhance prediction alignment across views. Meanwhile, the latter utilizes minimum spanning trees to establish pixel affinity relationships and generate structure-aware soft pseudo labels for online semantic guidance. Next, we introduce asymmetric regional contrast learning. This approach fuses high-confidence predictions from the weak augmentation branch with scribbles to form augmented labels, and construct category prototypes in the representation space. Simultaneously, we prioritize contrastive constraints on challenging pixels in boundary-unlabeled regions. This strategy enhances intra-class cohesion and inter-class separation at the representation level, effectively reducing category confusion in low-contrast edges and complex backgrounds. Experiments conducted on the constructed graphene and MoS2 datasets demonstrate that our method TACoS achieves over 96% of fully supervised performance using less than 0.6% annotated data. Furthermore, it exhibits superior structural coherence and boundary stability in scenarios with weakly contrasting edges and complex backgrounds, providing an efficient and scalable solution for automated high-throughput screening of 2D material flakes.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07169
PDF
-
NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction
2026-07-08 09:02:18Xiangyu Sun, Liu Liu, Seungkwon Yang, Jingbing Han, Seungtae Nam, Zhizhong Su, Eunbyung ParkAbstract
Pose-Free Feed-forward 3D Gaussian Splatting (3DGS) has recently emerged as a powerful paradigm for fast scene reconstruction. However, its performance degrades significantly in long image sequences due to cumulative camera pose estimation drift, which propagates errors into geometric modeling and severely limits rendering fidelity. In this work, we revisit the long-sequence bottleneck and identify pose drift as the primary factor restricting reconstruction quality. Furthermore, while SfM-based pseudo ground-truth poses introduce sensor noise, purely rendering-based supervision often leads to optimization instability and local minima due to the entangled optimization of geometry and pose. To address the challenges, we propose a synergistic pose-free framework that explicitly couples geometry and appearance via a Raymap-Guided Coupling Module (RGC). Concretely, we anchor Gaussian centers to raymap-induced geometry and jointly optimize RGB reconstruction, raymap consistency, and camera regularization under a unified objective, yielding a bidirectional feedback loop: stronger geometry improves rendering, and appearance supervision in turn refines geometry and pose. To further stabilize learning across wide temporal ranges, we introduce a Dual-Frequency Viewpoint Scheduling strategy that combines easy-to-hard interval expansion with replay of short-interval pairs. Extensive experiments across in-domain and cross-domain datasets show consistent gains in both rendering and pose estimation, with notably improved robustness on long sequences. Ablation studies validate our central insight: explicitly designed geometry-appearance synergy is the key to scalable and drift-robust pose-free feed-forward 3D reconstruction.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07168
PDF
-
From Text to Parameters: Predicting Item Parameters from Embedding Regularization with Reliability and Design Ceilings
2026-07-08 08:34:46Shi-Ting Chen, Jinsong ChenAbstract
Newly developed items must ordinarily be field tested before their psychometric properties are known, creating a cold start problem for item calibration. Predicting item parameters from features is a long standing measurement problem dating back to the Linear Logistic Test Model; modern text embeddings now automate the design matrices traditionally specified by hand. We propose an evaluation framework combining regularized regression on item text embeddings, repeated cross validated R squared reported with its resampling standard deviation, and two performance upper bounds: a reliability ceiling derived from parameter standard errors, and a design ceiling derived from simulation based power calibration. Applying this framework to a mathematics item bank (EEDI) and a medical licensure benchmark (BEA 2024), we find that item difficulty is highly predictable from text (repeated cross validated R squared = 0.53, or about 57% of its reliability ceiling), whereas discrimination and pseudo guessing appear less predictable. However, evaluating these results against our ceilings reveals that this apparent hierarchy stems from target reliability rather than text signal strength: text uniformly recovers 57 to 63% of the reliable variance across difficulty targets, whereas the 3PL pseudo guessing parameter has a reliability ceiling near zero, making it an unviable target at current precision. On BEA, embedding based regression matches leaderboard RMSE despite explaining almost no variance, highlighting the critical need for scale free metrics and explicit ceilings in benchmarking. Finally, we show that a single train and test split can inflate apparent accuracy by 0.1 to 0.15 in R squared, underscoring the necessity of repeated cross validation for calibration support applications and future benchmark construction.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07141
PDF
-
Hybrid Least Squares/Gradient Descent Methods for MIONets
2026-07-08 03:59:25Jun Choi, Chang-Ock Lee, Minam MoonAbstract
In this paper, we propose an efficient hybrid least squares/gradient descent (LSGD) method for MIONets to accelerate training. This method generalizes the LSGD method for DeepONets. Since MIONet is the sum of the entrywise product of multiple branch networks and a trunk network, it can be viewed as a multilinear function with respect to the last layer parameters of each branch network. These sets of parameters can be optimized using the alternating least squares method, where we solve the LS system for a single branch network in turn. To handle the large-sized system matrix, we introduce Kronecker and Khatri-Rao products and tensor permutation matrices to factor the large matrix into small ones. Our method is compatible with a general type of $L^2$ loss with regularization terms for the last layer parameters of each branch, where linear operators can be applied to the MIONet output in each loss term.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06976
PDF
-
Imputation Meets Clustering: Exploiting Latent Subgroup Structure for Missing Data Recovery
2026-07-08 02:50:02Chuyao Zhang, E Li, Taochen Chen, Yiqun Zhang, Yuzhu Ji, Shuping Zhao, Peng Liu, Yiu-ming CheungAbstract
Missing data is prevalent in practical applications, making effective imputation an essential preprocessing step for downstream analysis. Real-world datasets often exhibit complex latent structures composed of multiple subgroups with distinct distributions. However, existing methods often overlook such population heterogeneity. Without explicit structural guidance, these methods tend to produce generic estimates that blur subgroup boundaries and lack instance-level fidelity. While incorporating subgroup information offers a remedy, it faces a circular dependency: reliable subgroup identification requires complete data, while data completion is the imputation objective itself. To resolve this, we propose CAGI (Cluster-Aware Generative Imputation), a framework that reformulates clustering and imputation as a mutually reinforcing co-optimization process. CAGI employs a ``Partition-Guide-Restore'' strategy where dynamic cluster assignments act as local priors to condition a Generative Adversarial Network. An iterative feedback loop is established to progressively refine both cluster structures and imputed values toward faithful subgroup distributions. To ensure distributional stability, CAGI further employs a multi-level optimization objective combining instance-level reconstruction with distribution-level regularization. Extensive experiments on 14 benchmark datasets with 15 representative baselines demonstrate the superiority of CAGI. The source code is available at: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2607.06930
PDF
-
Bi-PT: Bidirectional Cross-Attention Point Transformers for Four-Chamber Heart Reconstruction from Sparse Cardiac MRI Data
2026-07-08 02:37:55Chenchuhui Hu, Shaoming Pan, Leon Axel, Meng YeAbstract
We propose Bi-PT, a pipeline for reconstructing 3D four-chamber human heart meshes from clinical sparsely sampled cardiac magnetic resonance imaging (CMR) data. This work addresses the error-prone generation of 3D cardiac shape from a sparse point cloud (SPC) extracted from 2D long-axis and short-axis views used in routine clinical CMR protocols. Bi-PT enables accurate inference of the four-chamber heart mesh from the SPC by learning robust point features via bidirectional point cross-attention between an atlas and the SPC, together with per-point semantic labels that improve correspondence estimation. We formulate the deformation field as a Neural Ordinary Differential Equation (NODE) parameterized by a per-point affine transformation and translation to deform the atlas toward the target heart shape. By learning such a NODE, we can guarantee the deformation field to be a locally affine diffeomorphic deformation. We also integrate a semantic label loss into the Chamfer distance to encourage label-consistent correspondences and add a smoothness regularization to stabilize and improve the learning of the deformation field. Extensive experiments demonstrate that Bi-PT achieves accurate and robust performance compared to baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06923
PDF
-
RSF-GLLM: Bridging the Semantic Gap in Multi-Hop Knowledge Graph QA via Recurrent Soft-Flow and Decoupled LLM Generation
2026-07-07 17:32:36Sambaran Bandyopadhyay, Ananth MuppidiAbstract
Multi-hop Question Answering over Knowledge Graphs faces a critical challenge: traditional retrieve-then-read pipelines break differentiability, preventing the retriever from learning to bridge the semantic gap where intermediate nodes lack lexical overlap with the query. To address this, we propose RSF-GLLM, a framework decoupling differentiable graph reasoning from answer generation. Our Recurrent Soft-Flow (RSF) module employs a GRU-guided query updater to propagate continuous relevance scores, utilizing a dynamic gating mechanism to traverse semantically dissimilar bridge nodes via structural cues. We introduce flow sparsity regularization to theoretically guarantee convergence from soft probabilities to discrete reasoning paths. These paths are extracted and textualized to fine-tune a Large Language Model (LLM), ensuring generation is grounded in factual topology. Experiments on WebQSP and CWQ demonstrate that RSF-GLLM achieves competitive performance with superior inference efficiency compared to LLM based computationally expensive approaches.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06527
PDF
-
Prompt-Adapter Context Routing for Parameter-Efficient Multi-Shot Long Video Extrapolation
2026-07-07 16:41:25Anna C\'ordoba, Adam Puente Tercero, Nerea Angulo Hijo, Mar Linares Tercero, Julia Barrientos, Ainhoa Miranda, Jes\'us OliveraAbstract
We present PACR-Video, a parameter-efficient framework for multi-shot long video extrapolation that preserves recurring entities, scene structure, visual style, and causal progression without full generator fine-tuning. PACR-Video keeps a text-to-video diffusion transformer frozen and augments it with low-rank temporal adapters conditioned by learned shot-role prompt tokens. To maintain long-horizon coherence, it builds a recursive prompt bank that stores compact entity, location, action, and style prompts from previous shots, then routes them through adapter gates according to predicted narrative dependencies. A Shot-Local/Story-Global tuning objective combines next-shot reconstruction, cross-shot identity contrast, and prompt sparsity regularization, while an adapter composition schedule balances early-shot visual consistency with later-shot event progression and viewpoint change. Across six multi-shot and long-video benchmarks, PACR-Video outperforms text-to-video, tuning-based, memory-augmented, streaming, and recursive-context baselines on distributional quality, semantic alignment, identity consistency, temporal smoothness, motion stability, transition coherence, and human preference. These results show that compact prompt routing and lightweight temporal adaptation provide sufficient controllable capacity for stable long video extrapolation.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06481
PDF
-
Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization
2026-07-07 05:34:25Kaishen Wang, Tong Zheng, Xuehao Cui, Ruibo Chen, Tianyi Xiong, Heng HuangAbstract
Large reasoning models (LRMs) improve language model capabilities by generating explicit thinking traces before final answers. In factuality-oriented question answering (QA), such thinking often improves overall performance by helping the model recover relevant knowledge and refine its answers. However, we find that this benefit is not uniform at the instance level: explicit thinking can also overturn correct non-thinking answers and lead to factual drift. We refer to this failure mode as \emph{thinking-induced hallucination}. To explain this phenomenon, we formulate explicit thinking in factuality QA as a thinking residual over the model's direct-answer tendency, which can either recover missing knowledge or introduce unsupported associations. Based on this formulation, we propose MARGO, \underline{\textit{M}}ixed-Mode \underline{\textit{A}}dvantage \underline{\textit{R}}egularization for \underline{\textit{G}}rounded \underline{\textit{O}}ptimization, a reinforcement learning framework that uses non-thinking rollouts as same-model references in advantage estimation. By constructing mixed-mode rollout groups with both thinking and non-thinking trajectories, MARGO evaluates whether explicit thinking adds factual value beyond direct answering, thereby suppressing hallucination-prone thinking while preserving beneficial thinking behaviors. Experiments across multiple factuality-oriented QA benchmarks demonstrate that MARGO improves factual reliability over strong baselines, while evaluations on mathematical benchmarks show that it preserves general reasoning ability.
Abstract (translated)
URL
https://arxiv.org/abs/2607.05861
PDF
-
$mathbf{lambda}$-VAE: Variance Equalization for Posterior Collapse
Abstract
Variational Autoencoders (VAEs) frequently suffer from posterior collapse, a failure mode in which the approximate posterior converges to the prior, rendering the latent code uninformative. Despite extensive research, a unified account of why collapse occurs has remained an open question. We identify and formalize two logically independent but coupled causes. \emph{Gradient imbalance} occurs when the decoder's reconstruction signal vanishes faster than the $\mathbb{KL}$ regularization pressure as the posterior widens. \emph{Information gap} occurs when the stochastic sampling step discards a substantial fraction of the encoder's computed representation, attenuating decoder sensitivity and making collapse inexpensive. Both causes share the same collapse trajectory, and we show that the information gap is algebraically equivalent to mismatch between the aggregate posterior and the prior, unifying two pathologies. Subsequently, we introduce $\lambda$-VAE, which resolves both causes through a single modification to the reparameterization step: the sampling noise is scaled by per-dimension exponent, while the $\mathbb{KL}$ penalty retains the original posterior variance. This asymmetry shifts the stable training attractor away from the degenerate collapsed state, driving all latent dimensions toward the same equilibrium -- a mechanism we term \emph{variance equalization}. A closed-form optimal exponent per dimension follows from a net information gain objective, with a single hyperparameter controlling the reconstruction-generation tradeoff. We validate on standard benchmarks (Binary MNIST, Binary Omniglot, CIFAR-10, CelebA-64), showing consistent reductions in collapsed dimensions, information capacity gains of up to $2.8\times$ nats, and reconstruction quality improvements of up to $+0.33$ BPD.
Abstract (translated)
URL
https://arxiv.org/abs/2607.05531
PDF
-
Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models
2026-07-06 15:16:10Raj Jaiswal, Dhruv Jain, Rishabh Dhawan, Sree Krishna Uppalapati, Shin'ichi Satoh, Tanuja Ganu, Rajiv Ratn ShahAbstract
Physics reasoning fails structurally in small language models: an error at any step propagates forward, corrupting every inference that follows. Limited domain knowledge, hallucination under multi-step derivation, and distributional sensitivity compound this failure. We propose a step-level reward framework that identifies the first reasoning error, generates targeted structured feedback, and trains the model to revise its solution via policy gradient with KL regularization, without exposing it to ground truth solutions as generation targets. Unlike annotation-dependent step-level methods, no preference data construction is required and the external verifier operates exclusively at training time. Across five physics benchmarks, our framework delivers accuracy gains of 17-20% over CoT prompting and 10-16% over the strongest baseline, reduces calculation errors from 56.9% to 23.5%, and reduces miscomprehension errors from 22.3% to 12.0% in the best observed cases. Conceptual errors reduce from 89.7% to 68.7%, yet persist as the hardest failure mode across all conditions.
Abstract (translated)
URL
https://arxiv.org/abs/2607.05199
PDF
