探索全球前沿学术脉络

01.

arXiv (CS.LG) 2026-06-16 DOI: arXiv:2508.03867

Constraining the outputs of ReLU neural networks

作者:

Yulia Alexandr↗Guido Mont\'ufar↗

arXiv:2508.03867v2 Announce Type: replace-cross Abstract: We introduce a class of algebraic varieties naturally associated with ReLU neural networks, arising from the piecewise linear structure of their outputs across activation regions in input space, and the piecewise multilinear structure in parameter space. By analyzing the rank constraints on the network outputs within each activation region, we derive polynomial equations that characterize the functions representable by the network. We further investigate conditions under which these varieties attain their expected dimension, providing insight into the expressive and structural properties of ReLU networks.

阅读与讨论 → 访问原文 →

02.

arXiv (CS.CL) 2026-06-11 DOI: arXiv:2605.04893

Self-Attention as Transport: Limits of Symmetric Spectral Diagnostics

作者:

Dominik Dahlem↗Diego Maniloff↗Mac Misiura↗

When a language model processes a hallucinated response, its attention routing tends to fail in one of two shapes: over-concentrating on a narrow set of positions, or spreading so diffusely that relevance is diluted, and the shape of the failure carries diagnostic signal. We study these shapes as a diagnostic characterization, computed from attention matrices under forced scoring of benchmark-labeled responses rather than during live generation. A widely used family of spectral methods analyzes the symmetric component of the degree-normalized attention operator, which governs transport capacity; we prove that every transpose-invariant spectral diagnostic of this operator is structurally orientation-blind (it cannot distinguish an operator from its transpose, and therefore cannot detect information-flow direction), with a converse to the blindness theorem bounding any Lipschitz diagnostic's transpose sensitivity by the asymmetry coefficient $G$. Pairing this with a closed-form bipartite-Cheeger landscape for canonical causal architectures, we show that uniform causal attention satisfies an $n$-independent floor $\phi \ge 1/5$, while window attention pierces the floor as $O(w/n)$; failure modes are shape-different, not just value-different. This floor is an idealized-architecture benchmark, not an empirical attractor: the fraction of real attention heads that pierce it is itself an architectural signature. The resulting two-axis diagnostic ($\phi$ for capacity, $G$ for direction) yields a falsifiable polarity prediction: bottleneck- and diffuse-dominated benchmarks should exhibit opposite polarity. Under length-controlled evaluation, transport features retain interpretable signal (0.62-0.84 LC-AUROC) across the tested decoder-only, encoder-only, and encoder-decoder models, with polarity reversing as predicted between HaluEval and MedHallu.

阅读与讨论 → 访问原文 →

03.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2606.16415

Posterior Twins: Distributional Behavioral Simulation for Enterprise Decisions

作者:

Ankit Das↗

arXiv:2606.16415v1 Announce Type: new Abstract: Enterprise behavioral simulation requires more than producing a plausible response. Many decisions depend on the shape of a population under a proposed action: which segments accept, defect, hesitate, or move into risk-sensitive states. This paper introduces Posterior Twins, a memory-grounded digital-twin approach that represents likely behavior as an updated distribution under a specific decision context. We evaluate a family of Twinning Labs behavioral-model operating points on a 226-example held-out behavioral-response benchmark and report both modal accuracy and Wasserstein-1 distance. The results show that modal accuracy and distributional fidelity identify different operating regimes. TL-Twin Alpha achieves the lowest observed Wasserstein-1 distance in the reported result set ($W_1 = 1.16$), while TL-Twin Delta and TL-Twin Gamma provide balanced operating points near the modal-accuracy frontier. The paper frames these results as a systems result: governed memory, behavioral model routing, scenario orchestration, distributional aggregation, and auditability are necessary for turning simulated behavior into reusable enterprise decision evidence.

阅读与讨论 → 访问原文 →

04.

arXiv (CS.AI) 2026-06-19 DOI: arXiv:2601.02149

AI-enhanced tuning of quantum dot Hamiltonians toward Majorana modes

作者:

Mateusz Krawczyk↗Jaros{\l}aw Paw{\l}owski↗

arXiv:2601.02149v4 Announce Type: replace-cross Abstract: We propose a neural network-based model capable of learning the broad landscape of working regimes in quantum dot simulators, and using this knowledge to autotune these devices - based on transport measurements - toward obtaining Majorana modes in the structure. The model is trained in an unsupervised manner on synthetic data in the form of conductance maps, using a physics-informed loss that incorporates key properties of Majorana zero modes. We show that, with appropriate training, a deep vision-transformer network can efficiently memorize relation between Hamiltonian parameters and structures on conductance maps and use it to propose parameters update for a quantum dot chain that drive the system toward topological phase. Starting from a broad range of initial detunings in parameter space, a single update step is sufficient to generate nontrivial zero modes. Moreover, by enabling an iterative tuning procedure - where the system acquires updated conductance maps at each step - we demonstrate that the method can address a much larger region of the parameter space.

阅读与讨论 → 访问原文 →

05.

arXiv (CS.LG) 2026-06-16 DOI: arXiv:2605.31027

Multi-Scale Separable Fourier Neural Networks for Solving High-Frequency PDEs

作者:

Qihong Yang↗Qiaolin He↗

arXiv:2605.31027v2 Announce Type: replace Abstract: We propose a novel neural network architecture, termed Multi-Scale Separable Fourier Neural Networks (MS-SFNN), for the accurate and efficient solution of linear and nonlinear high-frequency partial differential equations (PDEs). MS-SFNN exploits a separable representation: given a $d$-dimensional input, it employs $d$ independent subnetworks – each acting on a single coordinate – and constructs basis functions via element-wise multiplication of their outputs. The PDE solution is approximated as a linear combination of these basis functions, with coefficients determined by least squares. Critically, all network weights and biases are randomly initialized once, from a uniform distribution with unit variance, and remain fixed thereafter. To enhance expressivity, a tunable scaling factor is introduced in each subnetwork to modulate the frequency content of the resulting basis functions. Fourier features are explicitly embedded through cosine activations, endowing the method with strong spectral approximation capabilities. To mitigate the memory bottleneck associated with dense collocation in high-frequency or three-dimensional problems, we replace automatic differentiation with analytically derived basis function derivatives and develop a memory-efficient batched QR decomposition algorithm for solving large-scale least-squares systems. Numerical experiments demonstrate that MS-SFNN achieves unprecedented accuracy across a range of challenging PDEs, significantly outperforming state-of-the-art methods such as Physics-Informed Neural Networks (PINN) and Separated-Variable Spectral Neural Networks (SV-SNN).

阅读与讨论 → 访问原文 →

06.

arXiv (CS.LG) 2026-06-15 DOI: arXiv:2606.14139

Decoupled Latent Optimization of Diffusion Models for Full Waveform Inversion

作者:

Chen Min↗Zheng Ma↗

arXiv:2606.14139v1 Announce Type: new Abstract: Full waveform inversion (FWI) recovers subsurface velocity from seismic recordings by solving a severely ill-posed, nonconvex PDE-constrained optimization. Classical regularizers stabilize the inversion but fail to reproduce realistic geological structures; recent diffusion-prior methods improve realism at the cost of a fragile trade-off between data fidelity and prior consistency. We propose Decoupled Latent Optimization (DLO), which relaxes the standard latent-optimization formulation into a quadratic-penalty objective over an auxiliary physical variable and a latent variable. The data-fidelity gradient acts in physical space, the diffusion sampler contributes only through a decoded prior sample, and the standard smoothed-velocity initialization of classical FWI is preserved. On the OpenFWI benchmark, DLO outperforms classical regularizers and existing diffusion-based methods under clean, noisy, and missing-trace acquisitions. The prior, trained on 70*70 OpenFWI models, transfers directly to the Marmousi and Overthrust benchmarks, where DLO recovers intricate fault structures and remains robust to initialization smoothing and measurement noise.

阅读与讨论 → 访问原文 →

07.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2606.16152

The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning

作者:

Haolong Qian↗Xianliang Yang↗Yinuo ma↗Lirong Che↗Feng Lu↗Ye Guo↗Lei Song↗Jiang Bian↗Chun Yuan↗

arXiv:2606.16152v1 Announce Type: new Abstract: Knowledge distillation from powerful reasoning models is widely used to improve Small Language Models (SLMs) on mathematical reasoning, often assuming that traces with higher reward model scores provide more useful supervision. We identify a counterintuitive Quality-Utility Paradox in mathematical reasoning distillation. Data refined or synthesized by a stronger Oracle obtains higher perceived quality according to reward models, yet consistently underperforms traces generated by the SLM itself and selected through rejection sampling across Qwen2.5, LLaMA-3, and DeepSeek families. Our analysis shows that Oracle refinement couples logical repair with distributional drift away from the SLM's native reasoning distribution. This drift increases the learner's adaptation cost and can outweigh the benefit of improved reasoning logic. To test this mechanism, we introduce Style-Aligned Refinement, which preserves the native trajectory of the SLM while retaining logical repair from the Oracle. This intervention lowers adaptation cost and restores downstream utility. These findings suggest that effective mathematical reasoning distillation should jointly optimize perceived solution quality and learner-data compatibility, rather than relying solely on reward-model scores. The datasets and code are available at https://github.com/Dracoqhl/Quality-Utility-Paradox.

阅读与讨论 → 访问原文 →

08.

medRxiv (Medicine) 2026-06-18 DOI: HASH:6a31e9bebe50cbbe383b90b06e4a0a10

Cardiac rhythm development: A wearable device index of risk for physical and mental illness in adolescence

作者:

Giampetruzzi↗Kircanski↗Pine↗D. S↗Gotlib↗I. H↗

Objective. The autonomic nervous system, which regulates cardiac rhythm, undergoes pronounced maturation across adolescence. How cardiac rhythm develops over this period, however, and whether individual differences in its development forecast mental and physical illness, remain open questions. We used three waves of Fitbit data from the Adolescent Brain Cognitive Development (ABCD) Study to characterize the developmental trajectory of the cardiac rhythm and to test whether variation in that trajectory predicts onset of psychopathology and cardiometabolic disease. Methods. 8,301 adolescents contributed 242,811 valid Fitbit wear days across Waves 2 (Mage=12), 4 (Mage=14), and 6 (Mage=16). Cosinor mixed-effects models yielded three rhythm parameters per session: mesor (24-hour mean), amplitude (diurnal swing), and acrophase (peak timing). We first characterized age- and sex-specific trajectories, cross-wave stability, and factors shaping the rhythm. We then used parallel-process latent growth models to test whether within-person changes in rhythm tracked symptom trajectories, and hierarchical logistic models to test whether rhythm parameters predicted the first clinical onset of psychopathology and of obesity and hypertension. Results. The cardiac rhythm changed substantially across adolescence: mesor decreased, amplitude flattened, and acrophase shifted later. Within-person change in the rhythm tracked change in blood pressure, BMI, and trajectories of depression and ADHD symptoms. Higher mesor predicted incident onset of all five outcomes controlling for demographics, baseline symptoms, and behavior (ORs 1.36-1.54); amplitude, acrophase, and rhythm instability conferred additional risk. Conclusions. The 24-hour cardiac rhythm is a passively measurable substrate of adolescent autonomic development that indexes transdiagnostic risk for psychiatric and cardiometabolic illness.

阅读与讨论 → 访问原文 →

09.

arXiv (CS.CL) 2026-06-15 DOI: arXiv:2606.14122

Beyond Perplexity: UTF-8 Validity in Byte-aware Language Models

作者:

Sangwhan Moon↗Daisuke Oba↗Youmi Ma↗Tatsuya Hiraoka↗Naoaki Okazaki↗

Byte-level tokenization enables language models to handle any Unicode input, but models can generate invalid UTF-8 sequences when encountering rare or unseen characters. We investigate the relationship between training scale and UTF-8 generation reliability with a 355M parameter model trained on 80B tokens from a balanced multilingual corpus of English, Japanese, Korean, and Chinese. We introduce multiple evaluation protocols that isolate UTF-8 structural validity from language modeling. UTF-8 validity convergence lags perplexity by a roughly a factor of two: perplexity stabilizes after 2.1B tokens, but UTF-8 validity requires 4.2B tokens. In context-free generation, rare characters achieve higher structural validity than common characters, suggesting over-specialization of frequent character representations. Through experiments, we observed that reliable UTF-8 generation is a distinct capability requiring evaluation beyond perplexity.

阅读与讨论 → 访问原文 →

10.

arXiv (CS.LG) 2026-06-12 DOI: arXiv:2603.11249

Differentiable Thermodynamic Phase-Equilibria for Machine Learning

作者:

Karim K. Ben Hicham↗Moreno Ascani↗Jan G. Rittig↗Alexander Mitsos↗

arXiv:2603.11249v3 Announce Type: replace Abstract: Accurate prediction of phase equilibria remains a central challenge in chemical engineering. Physics-consistent machine learning methods that incorporate thermodynamic structure into neural networks have recently shown strong performance for activity-coefficient modeling. However, extending such approaches to equilibrium data arising from an extremum principle, such as liquid-liquid equilibria, remains difficult. Here we present DISCOMAX, a differentiable algorithm for phase-equilibrium calculation that guarantees thermodynamic consistency at both training and inference, only subject to a user-specified discretization. The method combines discrete enumeration of feasible phase states with masked softmax aggregation in the backward pass, with the propagation of the true equilibrium state in the forward pass, using a straight-through gradient estimator to enable physics-consistent end-to-end learning of neural \gls{gE}-models. We show that this approach bears analogy to statistical thermodynamics, and we evaluate it on binary liquid-liquid equilibrium data where it outperforms existing surrogate-based methods, while offering a general framework for learning from different kinds of equilibrium data.

阅读与讨论 → 访问原文 →

11.

arXiv (CS.AI) 2026-06-17 DOI: arXiv:2606.17648

From Brewing to Resolution: Tracing the Internal Lifecycle of Code Reasoning in LLMs

作者:

Siyue Chen↗Yifu Guo↗Yuquan Lu↗Zishan Xu↗Jiaye Lin↗Jianbo Lin↗Siyu Zhang↗Cheng Yang↗Junxin Li↗Yujia Li↗Yu Huo↗Ruixuan Wang↗…

arXiv:2606.17648v1 Announce Type: new Abstract: Standard accuracy metrics cannot explain why LLMs handle variable tracking but fail on semantically equivalent loops. We study an internal lifecycle of code reasoning in which models first brew the answer, making it linearly recoverable many layers before it becomes self-decodable, and then diverge into one of four resolution outcomes: Resolved, Overprocessed, Misresolved, or Unresolved. Understanding this lifecycle matters because similar task accuracies can mask fundamentally different failure modes that surface-level evaluation cannot detect. We introduce a dual diagnostic framework pairing layer-wise linear probing with Context-Stripped Decoding (CSD) and apply it to six code-reasoning task families across 16 models spanning Qwen, Llama, and DeepSeek architectures. All four outcomes carry substantial mass in every task family: overall Resolved is only 41.5%, with multiple tasks below 30%. Controlled sweeps over structure, depth, and operators expose task-specific failure bottlenecks: Function Call Resolved plunges from 61.1% to 2.5% as call depth increases from one to three. Across architectures and scales, the brewing scaffold remains stable, with normalized brewing duration 24-42% across all 16 models, while resolution success varies with capability. This indicates that the scaffold is a stable empirical regularity across the tested decoder-only Transformer families, whereas resolution success covaries with capability, scale, and training. Code: https://github.com/euyis1019/llm-brewing

阅读与讨论 → 访问原文 →

12.

arXiv (CS.CV) 2026-06-16 DOI: arXiv:2606.15681

3D Consistency Optimization for Self-Supervised Monocular Video Depth Estimation

作者:

Yuanye Liu↗Ke Zhang↗Junzhe Jiang↗Li Zhang↗Vishal Patel↗Xiahai Zhuang↗

Reliable monocular video depth estimation is crucial for downstream 3D reasoning and embodied AI in endoscopic navigation. However, existing self-supervised approaches typically treat video frames independently or rely on weak temporal regularization. These methods, lacking a holistic perception of the underlying 3D scene, inevitably suffer from geometrically inconsistent predictions and severe cross-frame drift. To address these limitations, we introduce a new paradigm that recasts sequential video depth estimation as an unconstrained multi-view 3D reconstruction problem, enabling full exploitation of the powerful geometric priors embedded in recent 3D foundation models. The core of our approach is a 3D consistency optimization framework driven by three constraints: image-level photometric rendering, explicit world-coordinate geometric alignment, and multi-scale temporal gradient consistency. Such unified optimization elegantly anchors isolated frames to a globally coherent 3D structure. Our method has been validated in both the self-supervised training scenarios and challenging zero-shot clinical environments. Results show that the proposed approach achieves state-of-the-art spatial accuracy, outperforming the frame-based, video-based depth estimators and the multi-view 3D reconstruction baselines.

阅读与讨论 → 访问原文 →

13.

arXiv (CS.CV) 2026-06-16 DOI: arXiv:2602.12279

UniT: Unified Multimodal Chain-of-Thought Test-time Scaling

作者:

Leon Liangyu Chen↗Haoyu Ma↗Zhipeng Fan↗Ziqi Huang↗Animesh Sinha↗Xiaoliang Dai↗Jialiang Wang↗Zecheng He↗Jianwei Yang↗Chunyuan Li↗Junzhe Sun↗Chu Wang↗…

Unified models can handle both multimodal understanding and generation within a single architecture, yet they typically operate in a single pass without iteratively refining their outputs. Many multimodal tasks, especially those involving complex spatial compositions, multiple interacting objects, or evolving instructions, require decomposing instructions, verifying intermediate results, and making iterative corrections. While test-time scaling (TTS) has demonstrated that allocating additional inference compute for iterative reasoning substantially improves language model performance, extending this paradigm to unified multimodal models remains an open challenge. We introduce UniT, a framework for multimodal chain-of-thought test-time scaling that enables a single unified model to reason, verify, and refine across multiple rounds. UniT combines agentic data synthesis, unified model training, and flexible test-time inference to elicit cognitive behaviors including verification, subgoal decomposition, and content memory. Our key findings are: (1) unified models trained on short reasoning trajectories generalize to longer inference chains at test time; (2) sequential chain-of-thought reasoning provides a more scalable and compute-efficient TTS strategy than parallel sampling; (3) training on generation and editing trajectories improves out-of-distribution visual reasoning. These results establish multimodal test-time scaling as an effective paradigm for advancing both generation and understanding in unified models.

阅读与讨论 → 访问原文 →

14.

arXiv (CS.CL) 2026-06-12 DOI: arXiv:2606.13189

SICI: A Semantic-Pragmatic Complexity Index Reveals Regime Shifts in LLM Stance Detection

作者:

Fuqiang Niu↗Bowen Zhang↗

Prompt-based LLMs are increasingly used for stance detection, but harder examples are not always repaired by clearer instructions, reasoning prompts, retrieval, or debate. We introduce SICI (Stance Inference Complexity Index), a seven-dimensional diagnostic measure of the semantic-pragmatic burden imposed by a target–text pair. Across SemEval-2016 and VAST, SICI predicts LLM accuracy better than surface proxies and shows substantial cross-scorer reliability ($\alpha=0.771$). More importantly, LLM errors change regime as SICI increases: low-complexity examples invite over-attribution, especially Against predictions; intermediate examples form an unstable boundary; and high-complexity examples rapidly concentrate on None. This phase-transition-like structure persists across GPT-3.5, GPT-4o-mini, DeepSeek-V3, and GPT-4o, although stronger models move the boundaries. A 15-method intervention study further shows that prompting, retrieval, and debate often shift models along the attribution–abstention axis rather than removing the high-complexity bottleneck.

阅读与讨论 → 访问原文 →

15.

arXiv (CS.CL) 2026-06-16 DOI: arXiv:2606.16360

Tyler: Typed Latent Reasoning for Language Models – When to Think, What to Compute, and How Much to Allocate

作者:

Hanyu Lin↗Min Cai↗Jiawei Wen↗Haodi Zhang↗

Chain-of-thought (CoT) prompting improves reasoning in large language models (LLMs) by externalizing intermediate computation as discrete text tokens, but this textual interface also introduces redundancy and inference overhead. Latent reasoning offers a promising alternative by carrying part of the computation in continuous representations. However, existing methods typically predefine when latent computation is invoked and how it is allocated during decoding, leaving a key problem unresolved: when to invoke latent computation, what type of computation to perform, and how much budget to allocate. We propose Typed Latent Reasoning (Tyler), a typed and budget-aware framework for latent reasoning during autoregressive decoding. Tyler learns a policy that, at each decoding step, chooses between emitting a text token and switching to a latent computation module specialized for a particular reasoning function. Once invoked, an operator maps the current reasoning state into latent tokens that support global planning, local state updates, or reusable procedural abstraction. Across extensive experiments on three backbone LLMs, Tyler improves accuracy by up to 14.49 points over CoT and by up to 4.30 points over the strongest competing baseline. It further generalizes across diverse reasoning domains and achieves the best final-stage performance with the lowest forgetting.

阅读与讨论 → 访问原文 →

16.

medRxiv (Medicine) 2026-06-18 DOI: HASH:241755eb4546a393bea28600c499bfca

Empirical Validation and Predictive Utility of the Perinatal Grief Scale in Men after Perinatal Loss

作者:

Ravaldi↗Mosconi↗Nespoli↗Fumagalli↗Vannacci↗

Background. The Perinatal Grief Scale (PGS) is a widely used instrument for assessing grief following pregnancy loss, yet no study has validated it specifically in men despite documented use in several studies. This gap is critical given fathers' persistent underrepresentation in perinatal bereavement research and the absence of empirically supported screening thresholds for this population. Methods. This cross-sectional validation study used data from the OPALE project (Observatory on PerinatAL hEalth) conducted by the CiaoLapo Foundation in Italy. Among 276 fathers who experienced stillbirth or miscarriage, we examined criterion validity by testing the association between PGS scores and trauma-related symptomatology assessed via three validated instruments: the Revised Impact of Event Scale (RIES, n=103), National Stressful Events Survey Short Scale (NSESSS, n=95), and SCL-90 (n=173). We systematically tested multiple threshold combinations to identify optimal discriminative performance. Results. The PGS demonstrated excellent criterion validity. The optimal threshold (PGS >=92) showed sensitivity 81.0%, specificity 81.8%, and Youden's J index 0.628. Fathers scoring >=92 had 19.12 times the odds of high trauma symptoms (95% CI: 9.35 to 39.14, p

阅读与讨论 → 访问原文 →

17.

arXiv (CS.LG) 2026-06-15 DOI: arXiv:2606.14690

A Complexity Measure for Active Learning in Multi-group Mean Estimation

作者:

Abdellah Aznag↗Rachel Cummings↗Adam N. Elmachtoub↗

arXiv:2606.14690v1 Announce Type: new Abstract: We study a max-risk objective for active learning in a multi-group mean estimation $d$-armed bandits: a learner adaptively allocates a budget of $T$ samples across $d$ groups to minimize the worst-case uncertainty index $\max_{k\in[d]}\sigma_k^2/n_k$, where $\sigma_k$ is the standard deviation of the distribution of arm $d$, and $n_k$ is the number of times arm $d$ is sampled. We develop a local minimax framework and prove the first general lower bound for this objective, valid for any finite-variance hypothesis class. The bound separates difficulty into three orthogonal factors: a budget term, a heteroscedasticity index measuring how unevenly the uncertainty is spread across arms, and a model-dependent complexity measure, the Variance Local Curvature ($\mathrm{VLC}$), which captures how much information a local change of variance creates inside the hypothesis class. For smooth classes, the $\mathrm{VLC}$ is a reparametrization of a variance–Fisher information, with closed-form values for common families. Benchmarking against the strongest available upper bound shows near-optimality up to logarithmic factors in broad regimes, and pinpoints a systematic gap in highly heterogeneous instances. Our proof introduces two key ingredients: a loss-induced $\ell_1$ geometry on the decision space, and a representation-based instance generator that reduces hard-instance construction to an explicit random matrix calculation.

阅读与讨论 → 访问原文 →

18.

arXiv (CS.AI) 2026-06-18 DOI: arXiv:2606.18322

SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior

作者:

Mingyue Cui↗Linghui Shen↗Xingyi Yang↗

arXiv:2606.18322v1 Announce Type: cross Abstract: Sparse Autoencoders (SAEs) decompose residual-stream activations into interpretable features. Recent latent-space defenses increasingly rely on these decompositions, assuming that identified "unsafe" SAE features serve as actionable handles for monitoring and intervention. In this paradigm, clamping a specific harmful feature is expected to reliably prevent model misbehavior. However, we show that this success may hide a recoverable failure mode: the clamp may block one visible route to a behavior without eliminating the behavior itself. We formulate this vulnerability as post-intervention recovery, a constrained residual-space optimization problem. Starting from the post-intervention residual state, we optimize residual perturbations to recover the pre-intervention behavior while preserving the post-intervention values of the targeted SAE features. Even under a strong threat model where the intervention remains active throughout optimization and generation, recovery remains possible. To rule out that recovery simply undoes the intervention, we use encoder-orthogonal updates for single-layer interventions and the corresponding feature-map Jacobian in the cross-layer setting. Across TPP, unlearning, IOI, and refusal steering experiments, this stress test reveals recoverable behavior despite successful feature-level intervention. Especially in the safety-critical refusal-steering setting, we achieve a 95.8% recovery rate on valid samples while keeping defended-feature relative drift to 0.131, substantially below suffix-based baselines. A recovery-path attribution analysis further localizes this recovery to the SAE reconstruction residual, the component left unexplained by the SAE. These results expose a gap between feature-level control and behavioral completeness: SAE features can support causal intervention, but controlling them does not guarantee control over the underlying behavior.

阅读与讨论 → 访问原文 →

19.

arXiv (CS.LG) 2026-06-17 DOI: arXiv:2606.09049

Data augmented bootstrap: Unifying confidence interval construction by approximate invariance

作者:

Kevin Han Huang↗

arXiv:2606.09049v2 Announce Type: replace-cross Abstract: We propose the data augmented bootstrap (DAB), a framework for constructing confidence intervals from approximately invariant transformations of the data. As special cases, DAB recovers popular methods that rely on exact group symmetries, such as conformal prediction, wild bootstrap for Maximum Mean Discrepancy U-statistics and the recently proposed SymmPI. Meanwhile, DAB also recovers the classical bootstrap method, which exploits the dataset's approximate invariance under uniform sampling of data indices as the dataset size grows. For all DAB methods, we establish theoretical coverage results that interpolate between finite-sample and asymptotic guarantees according to the strength of the invariance, and without assuming a group structure. The approximate invariance is measured in the Kolmogorov distance and, for statistics that satisfy Gaussian universality, reduces to conditional mean and variance matching. This allows us to incorporate data augmentation (DA), a widely used machine learning heuristic based on approximate invariances, into known statistical methods. We empirically test the performance of incorporating DA into bootstrap, wild bootstrap and conformal prediction for simulated settings as well as for image, language and scientific data.

阅读与讨论 → 访问原文 →

20.

arXiv (CS.CL) 2026-06-16 DOI: arXiv:2606.15191

AmchiBias: Measuring Stereotypical Bias in Goan Identity Groups with a Minimal Pair Dataset in English and Konkani

作者:

Michelle Barbosa↗Sebastian Pad\'o↗Franziska Weeber↗

Socio-cultural stereotypical bias is an important consideration in the development and deployment of NLP systems. It is however often considered only at the national level, despite rich subnational socio-cultural structures. We present AmchiBias, the first benchmark for measuring socio-cultural stereotypical bias for the Indian state of Goa with its unique historically multicultural setting. It covers various Goan identity groups and comprises 313 minimal pairs across eight sociodemographic dimensions in both English and Devanagari Konkani. We then evaluate stereotypical bias in five multilingual encoder models on this benchmark. We find near-chance scores in Konkani, reflecting language incompetence for general multilingual models and a lack of Goan cultural competence for Indian language models. Queried in English, models with a stronger Indian language coverage show higher bias for pan-Indian groups than hyperlocal Goan groups. This suggests the English signal reflects pan-Indian pretraining associations rather than genuine Goan cultural knowledge. Our findings highlight a critical gap in low-resource multilingual NLP evaluation for hyperlocal community identities.

阅读与讨论 → 访问原文 →

21.

arXiv (CS.CL) 2026-06-12 DOI: arXiv:2606.12764

Detecting Functional Memorization in Code Language Models

作者:

Matthieu Meeus↗Anil Ramakrishna↗Matthew Grange↗Zheng Xu↗Luca Melis↗

Large language models (LLMs) are increasingly used to generate code at scale. Meanwhile, prior work has investigated whether training data may be recoverable from model outputs, by auditing the textual overlap between training examples and model generations. Code, however, can be functionally equivalent while textually dissimilar. In this work, we study functional memorization: extraction of functional logic beyond what verbatim metrics detect. We construct a counterfactual setup for Olmo-3-32B, comparing a midtrained model (exposed to target code) against a pretrained reference (not exposed). We prompt both models with Python function signatures and measure both textual and functional similarity (i.e., LLM-as-a-judge, execution-based). Our results show clear evidence of functional memorization, highlighting the need for auditing metrics that go beyond textual overlap.

阅读与讨论 → 访问原文 →

22.

arXiv (CS.LG) 2026-06-17 DOI: arXiv:2606.17417

A Closer Look at Failure Modes in Temporal Understanding of Large Audio-Language Models

作者:

Apoorva Kulkarni↗Kaousheik Jayakumar↗Sreyan Ghosh↗Sarah Wiegreffe↗Dinesh Manocha↗Ramani Duraiswami↗

arXiv:2606.17417v1 Announce Type: cross Abstract: Large Audio Language Models (LALMs) achieve strong performance on a variety of audio understanding tasks but continue to struggle with temporal reasoning, a fundamental capability central to human auditory perception. Understanding the causes of these failures remains challenging as existing benchmarks report performance gaps without probing underlying mechanisms. To address this, we introduce a benchmark with 1,657 questions across three foundational tasks designed specifically for mechanistic analysis. Examining model outputs across varying input settings (behavioral analysis) reveals that models often under-utilize audio when textual cues are available. We also provide the first causal mechanistic analysis of temporal reasoning failures in LALMs. Comparing attention upweighting against scaling, we find that redistributing attention across audio tokens is more effective than increasing audio attention. Targeting task-relevant tokens yields further gains. These findings suggest that modality imbalance alone cannot explain failures. Attention scaling at bottleneck layers improves accuracy from 55.9% to 59.1% without fine-tuning, demonstrating a promising direction for future work.

阅读与讨论 → 访问原文 →

23.

arXiv (CS.CV) 2026-06-16 DOI: arXiv:2606.02506

Question-Aware Evidence Ledgers for Video Relational Reasoning

作者:

Yilin Ou↗Mengshi Qi↗Huadong Ma↗

The VRR-QA challenge evaluates visual relational reasoning in videos, where answers often depend on implicit spatial relations, event boundaries, target identity, and dialogue context rather than a single salient frame. We present a test-time reasoning pipeline built around a strong GPT-5.5 video QA solver and a set of question-aware evidence ledgers. The initial solver answers each question from a uniform video representation, while routed ledgers are prompted to make the required targets, count units, reference frames, and temporal or spatial scope explicit for counting, spatial, endpoint, viewpoint, and dialogue reasoning. External tools such as open-vocabulary detection, depth cues, pair crops, ASR, and scene-graph ledgers are used only as evidence sources. A conservative gate keeps the current answer unless independent evidence uniquely supports a different option. The final evidence-gated pipeline achieves 92.95% overall accuracy and 93.79% macro accuracy on the challenge test split.

阅读与讨论 → 访问原文 →

24.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2606.15684

Multi-agent Framework for Time-Sensitive Complementary Collaboration in Minecraft

作者:

Juheon Yi↗Jinglu Wang↗Xiaoyi Zhang↗Yan Lu↗

arXiv:2606.15684v1 Announce Type: new Abstract: We present TickingCollabBench, a Minecraft-based multi-agent benchmark for a novel class of time-sensitive complementary collaboration tasks. Our benchmark reflects four core characteristics of real-world collaboration: agent heterogeneity, mandatory collaboration, dynamic environments, and strict real-time constraints with failure risks. To enable this, we develop the TickingCollab framework, which supports the generation of diverse dynamic environments and abstracts Minecraft's primitive APIs to enable declarative YAML task specifications for composing these events. Building on this, we design a feasibility-aware automated benchmark generation pipeline, where an LLM drafts structurally diverse task configurations and feasibility verifier filters out invalid ones using approximate constraints. Evaluations demonstrate that lang latency and inherent difficulty of coordinating under partial observability and agent heterogeneity cause LLMs to frequently fail under dynamic environments and fall significantly short of a global-knowledge oracle.

阅读与讨论 → 访问原文 →

25.

arXiv (CS.CV) 2026-06-19 DOI: arXiv:2606.19531

ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

作者:

Yuyang Zhang↗Wenyao Zhang↗Zekun Qi↗He Zhang↗Haitao Lin↗Jingbo Zhang↗Yao Mu↗Xiaokang Yang↗Wenjun Zeng↗Xin Jin↗

World Action Models (WAMs) commonly rely on video generation to bridge visual world modeling and robot control. However, video-based WAMs face three coupled limitations: dense multi-frame future tokens make inference costly, full video prediction spends capacity on action-irrelevant temporal and appearance details, and long-horizon future imagination may introduce errors that mislead action prediction. These issues raise a simple question: Does world action model really need video generation? We propose ImageWAM, a simple WAM framework that repurposes pretrained image editing models for robot action prediction. In contrast to video generation, image editing provides a better-matched prior: it only needs to model a target-frame transformation, focuses on action-relevant current-to-target visual differences, and grounds task instructions to localized visual changes through edit pretraining. In practice, ImageWAM does not decode the target frame at inference time; instead, it conditions a flow-matching action expert on the KV caches produced by image-editing denoising, using them as a compact world-action context. ImageWAM outperforms standard VLA baselines and matching competitive WAMs without additional policy pretraining across different simulator and real-world experiments. It also reduces FLOPs to 1/6 and latency to 1/4 of video-based WAMs. Attention analysis further shows that editing caches focus on task-relevant change regions, supporting image editing as an effective alternative to video-based world-action modeling.

阅读与讨论 → 访问原文 →