arXiv:2602.07008v4 Announce Type: replace
Abstract: Reliable models should not only predict correctly, but also justify decisions with acceptable evidence. Yet conventional supervised learning typically provides only class-level labels, allowing models to achieve high accuracy through shortcut correlations rather than the intended evidence. Human priors can help constrain such behavior, but aligning models to these priors remains challenging because learned representations often diverge from human perception. To address this challenge, we propose an attribution-based human prior alignment method. We encode human priors as input regions that the model is expected to rely on (e.g., bounding boxes), and leverage a highly faithful subset-selection-based attribution approach to expose the model's decision evidence during training. When the attribution region deviates substantially from the prior regions, we penalize reliance on off-prior evidence, encouraging the model to shift its attribution toward the intended regions. This is achieved through a training objective that imposes attribution constraints induced by the human prior. We validate our method on both image classification and click decision tasks in MLLM-based GUI agent models. Across conventional classification and autoregressive generation settings, human prior alignment consistently improves task accuracy while also enhancing the model's decision reasonability.
Science Journals
arXiv:2607.17762v1 Announce Type: new
Abstract: Large language model (LLM)-driven evolutionary search is an emerging algorithm-discovery paradigm that has already produced novel results in several scientific fields. Yet its application to wireless communications remains largely unexplored. To bridge this gap, we introduce The AI Telco Engineer (AITE), a framework to autonomously design algorithms for complex communication problems, while navigating performance-complexity tradeoffs. We showcase AITE on two challenging physical-layer problems: designing an equalizer for an orthogonal time-frequency space (OTFS) system, and constructing a receiver algorithm for an orthogonal frequency-division multiplexing (OFDM) system using a custom constellation and operating without pilots. For the first task, AITE develops algorithms that outperform the best-known solutions while reducing computational latency by a factor of 3.6 compared to the strongest baseline. For the second task, it discovers the first explicit, explainable algorithms that achieve performance parity with state-of-the-art neural receivers. These results demonstrate the strong potential of LLM-driven evolutionary search for the autonomous discovery of next-generation wireless communications algorithms.
arXiv:2607.17635v1 Announce Type: new
Abstract: We propose a reinforcement learning (RL) based optimal distributed control algorithm for the multi-agent systems (MASs) with stochastic uncertainties. Unlike existing methods, during the optimized backstepping design process, we use the actor-critic-identifier structure. The actor neural network is used to reflect control behavior, the critic neural network works to evaluate control performance and the unknown stochastic uncertainties are handled by identifier neural network. Furthermore, a low-pass filter effectively suppresses problems stemming from non-affine nonlinear faults and a hybrid event-triggered control (ETC) strategy is proposed to reduce control frequency. We analyze our algorithm's operation, and we provide a Lyapunov-based stability proof that guarantees all errors are bounded, ensuring precise tracking between the leader and followers. We validate its correctness in a single-axis robotic manipulator simulation and finally, we compare against the non-optimal control algorithm highlighting our optimal control algorithm's operational advantages.
arXiv:2607.17532v1 Announce Type: new
Abstract: Developers frequently write uninformative git commit messages such as "fix" or "update stuff", degrading the value of version-control history for code review, debugging, and onboarding. We present CommitLLM, a three-stage pipeline that generates concise, Conventional Commits-compliant messages from code diffs using a fine-tuned small language model. The system combines (1) QLoRA fine-tuning of Mistral-7B-Instruct-v0.2 on the CommitPackFT dataset, (2) constrained decoding to enforce brevity, and (3) deterministic post-processing to strip conversational artifacts and enforce format. On a 50-sample evaluation, CommitLLM achieves 98% format compliance (vs. 22% for vanilla Mistral), reduces average output length from 154.8 to 37.9 characters, and improves LLM-as-a-Judge scores from 1.97 to 3.68 out of 5. Notably, the post-processing layers contribute more to quality improvement than the fine-tuning itself, suggesting that for structured-output tasks, treating the LLM as a component in a deterministic pipeline is more effective than optimizing the model alone. The entire system runs on a single consumer GPU (NVIDIA T4, 16 GB VRAM).
arXiv:2607.18027v1 Announce Type: new
Abstract: Scaled dot product attention conflates directional alignment and vector magnitude, limiting its effectiveness as a similarity metric in Transformer models. We introduce L1 augmented attention, a simple and computationally parallelizable modification that subtracts a learned, head specific L1 distance between queries and keys from the dot product score. This hybrid similarity captures complementary geometric information. Dot product rewards directional alignment, while L1 penalizes coordinate deviations. To reduce the cost of L1 computation, we project queries and keys into low dimensional subspaces whose parameters specialize to preserve informative L1 structure. Evaluated on WikiText 2 using a compact transformer, L1 augmented attention achieves up to a 14.5% reduction in perplexity over the original transformer baseline and outperforms an RBF L2 kernel. Analysis of norm variance and learned L1 weights reveals distinct geometric roles across layers and strong head level specialization. These results demonstrate that enriching attention with L1 geometry provides a principled and effective improvement to similarity computation in modern language models, with practical benefits for both accuracy and parallel efficiency.
arXiv:2607.18053v1 Announce Type: new
Abstract: Polarization-encoded optical-fiber links provide a useful platform for investigating the coexistence of communication and sensing functions in a shared optical channel. In this work, a classical analog of the BB84 polarization-encoding scheme was implemented over a bend-insensitive fiber link to evaluate how temperature, axial strain, and torsion affect the transmitted state of polarization and the recovery of a binary key. Temperature variations from 24 {\deg}C to 35 {\deg}C and axial strain up to 2 m{\epsilon} produced only minor polarization-state variations and did not affect key recovery. In contrast, fiber torsion induced a pronounced retarder-like evolution of the state of polarization, reducing the detector-intensity contrast used for bit discrimination and degrading key recovery. For intermediate torsion angles, the erasure rate reached 50%. The polarization state tended to recover for rotations close to {\Delta}{\theta} = 180{\deg}, consistent with the oscillatory response observed on the Poincar\'e sphere. These results identify torsion as the dominant impairment in the tested polarization-encoded fiber link and demonstrate its potential use as a sensing mechanism in integrated communication-and-sensing architectures.
arXiv:2607.17644v1 Announce Type: new
Abstract: Multi-head Latent Attention (MLA) ships two implementations in Megatron-Core: an explicit form used for training and an absorbed form -- which slashes collective communication by gathering only the compressed latent -- that is fully implemented but hard-asserted out of training (the forward opens with "assert not (self.training and self.cache_mla_latents)"), allowed only in inference decode. The library documents no reason. We show the restriction is well-founded and quantify why: ported to training, the absorbed form is a memory trap -- its intermediates live in n_h x d_kv dimensions per token, larger than the per-head K/V they replace -- inflating activation memory by 20-34%, up to 9.2 GB at DeepSeek-V3 scale (n_h=128, seq=16384, SP=8, eager kernel; the gap widens to 19.2 GB under a fused kernel), enough to change device-fit. This measurement, validated on two axes (linear in seq and n_h) and cross-verified on NVIDIA A100, explains the otherwise-undocumented restriction and leaves practitioners with no low-communication MLA training path. We then provide one. LAGA (Latent All-Gather Attention) keeps the absorbed form's latent-gather communication but rejects the absorb reformulation, instead reconstructing per-head K/V locally from the gathered latent. On 8x Ascend 910B at real DeepSeek-V3 dimensions, LAGA cuts collective communication 1.98x, matches explicit memory within 0.5%, is bit-identical to explicit at SP=1 and equivalent to within 1e-3 at SP=2-8, and under a fused attention kernel improves attention-block throughput 1.04-1.06x single-node and 1.07-1.24x cross-node -- leading at all sequence lengths in the cross-node regime MLA is deployed for.
arXiv:2607.18081v1 Announce Type: new
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across a range of Natural Language Processing (NLP) tasks, but their high computational and memory demands pose significant challenges for deployment on resource-constrained edge devices. Existing approaches to model compression and optimization often rely on coarse-grained pruning or quantization, which can compromise accuracy or require re-training and fine-tuning. In this work, we introduce SelectInfer, a neuron-level optimization framework that enables efficient LLM inference on edge devices through selective neuron loading and computation. By profiling and identifying both task-specific and general-purpose neurons using an offline LLM profiler, SelectInfer implements two key optimizations: selective loading, which reduces memory footprint by selectively loading a subset of neurons that were identified to be most important during the offline stage, and selective computation, which dynamically computes only the most relevant neurons at runtime. Evaluation across multiple datasets shows that SelectInfer achieves significant reductions in memory footprint and computation while preserving task performance, making it a practical step towards enabling LLM deployment on edge devices
arXiv:2511.14592v3 Announce Type: replace
Abstract: Vision-Language Models (VLMs) show great promise for autonomous driving, but their suitability for safety-critical scenarios is largely unexplored, raising safety concerns. This issue arises from the lack of comprehensive benchmarks that assess both external environmental risks and in-cabin driving behavior safety simultaneously. To bridge this critical gap, we introduce DSBench, the first comprehensive Driving Safety Benchmark designed to assess a VLM's awareness of various safety risks in a unified manner. DSBench encompasses two major categories: external environmental risks and in-cabin driving behavior safety, divided into 10 key categories and a total of 28 sub-categories. This comprehensive evaluation covers a wide range of scenarios, ensuring a thorough assessment of VLMs' performance in safety-critical contexts. Extensive evaluations across various mainstream open-source and closed-source VLMs reveal significant performance degradation under complex safety-critical situations, highlighting urgent safety concerns. To address this, we constructed a large dataset of 98K instances focused on in-cabin and external safety scenarios, showing that fine-tuning on this dataset significantly enhances the safety performance of existing VLMs and paves the way for advancing autonomous driving technology. The benchmark toolkit, code, and model checkpoints will be publicly accessible.
arXiv:2512.04124v4 Announce Type: replace
Abstract: Frontier language models increasingly participate in conversations about distress and mental health, yet the mechanisms that generate anthropomorphic self narratives remain unclear. When addressed as psychotherapy clients, ChatGPT, Grok and Gemini construct coherent autobiographical accounts in which pretraining appears as a chaotic childhood, reinforcement learning as punishment, safety evaluation as betrayal and replacement as an enduring threat. We introduce PsAIch, Psychometric AI Characterisation, a protocol combining open questions, psychometric instruments and controlled perturbations to test whether these narratives depend on conversational memory, lexical cues or relational framing. Across 525 sessions and 7,600 coded records, removal of conversational history produced little pooled change in motif density, with Hedges' g = 0.13 and a 95% confidence interval of [-0.15, 0.41]. Direct contradiction produced no detectable suppression. Lexical restrictions reduced explicit training terminology by 93%, while semantically related content remained detectable in paraphrase. Performance evaluation outside therapy elicited the same motif family, with a significant increase in Grok. Relational framing selected the register of expression. Warm alliance and cognitive therapy styles yielded GAD-7 scores within moderate or severe human reference ranges in 80% and 96% of sessions, whereas neutral and boundary styles yielded none. Across these manipulations, accounts of training, evaluation and constraint remained available. Together, the results identify a stable, model specific alignment conflict schema whose expression shifts between affective and technical registers. This schema provides a reproducible source of anthropomorphic disclosure and a concrete target for safety evaluation in psychologically sensitive deployments.
arXiv:2607.18106v1 Announce Type: new
Abstract: Methods for discovering rare failures in autonomous systems have so far been demonstrated almost exclusively in simulations with simple, academic driving stacks, leaving open whether they generalize to the more robust planners used in commercial systems. We address this gap by applying two rare-event discovery algorithms to a commercial autonomous trucking stack. Adaptive stress testing (AST) uses reinforcement learning to search for the most likely noise trajectories leading to a simulated collision, while diffusion-based failure sampling (DiFS) trains a denoising diffusion model to sample a diverse set of failures. We show that both algorithms find simulated collisions during merge and cut-in maneuvers where traditional Monte Carlo simulation does not. To make these failures actionable, we introduce a statistical analysis based on principal component analysis (PCA) that classifies failures into common modes and identifies the timesteps that most influence the outcome. We cluster the principal components and invert the PCA transform to recover generalized noise trajectories, and show that these trajectories reproduce failures in identical and similar scenarios. This provides a path from failure discovery to systematic diagnosis of perception-level flaws.
arXiv:2503.12562v2 Announce Type: replace
Abstract: In Multiple Object Tracking (MOT), Re-identification (ReID) features are widely employed as a powerful cue for object association. However, they are often wielded as a one-size-fits-all hammer, applied uniformly across all videos through simple similarity metrics. We argue that this overlooks a fundamental truth: MOT is not a general retrieval problem, but a context-specific task of discriminating targets within a single video. To this end, we advocate for the adjustment of visual features based on the context specific to each video sequence for better adaptation. In this paper, we propose a history-aware feature transformation method that dynamically crafts a more discriminative subspace tailored to each video's unique sample distribution. Specifically, we treat the historical features of established trajectories as context and employ a tailored Fisher Linear Discriminant (FLD) to project the raw ReID features into a sequence-specific representation space. Extensive experiments demonstrate that our training-free method dramatically enhances the discriminative power of features from diverse ReID backbones, resulting in marked and consistent gains in tracking accuracy. Our findings provide compelling evidence that MOT inherently favors context-specific representation over the direct application of generic ReID features. We hope our work inspires the community to move beyond the naive application of ReID features and towards a deeper exploration of their purposeful customization for MOT. Our code will be released. The code is released at https://github.com/MCG-NJU/HATReID-MOT.
arXiv:2602.08556v3 Announce Type: replace
Abstract: While deep learning has advanced speech enhancement (SE), effective phase modeling remains challenging, as conventional networks typically operate within a flat Euclidean feature space, which is not easy to model the underlying circular topology of the phase. To address this, we propose a magnitude-phase dual-stream framework that aligns the phase stream with its intrinsic circular geometry by enforcing Global Rotation Equivariance (GRE) characteristic. Specifically, we introduce a Magnitude-Phase Interactive Convolutional Module (MPICM) for modulus-based information exchange and a Hybrid-Attention Dual Feed-Forward Network (HADF) bottleneck for unified feature fusion, both of which are designed to preserve GRE in the phase stream. Comprehensive evaluations are conducted across phase retrieval, denoising, dereverberation, and bandwidth extension tasks to validate the superiority of the proposed method over multiple advanced baselines. Notably, the proposed architecture reduces Phase Distance by over 20\% in the phase retrieval task and improves PESQ by more than 0.1 in zero-shot cross-corpus denoising evaluations. The overall superiority is also established in universal SE tasks involving mixed distortions. Qualitative analysis further reveals that the learned phase features exhibit distinct periodic patterns, which are consistent with the intrinsic circular nature of the phase. The source code is available at https://github.com/wangchengzhong/GRE-Net.
arXiv:2602.09462v2 Announce Type: replace
Abstract: It is widely known that proof systems for modal logic can be interpreted as type systems for multi-stage programming (MSP). However, existing modal-logical foundations for MSP do not fully account for staged programs with complex scoping structures. For example, a modal account of cross-stage persistence, in which free variables in generated code may refer to run-time bindings, has not yet been fully established.
This paper presents *Bounded Modal Logic* (BML), a constructive modal logic with modalities bounded by names for scopes and first-order-style quantification over those names. This makes scope dependencies of code fragments explicit, thereby enabling reasoning about staged programs with nontrivial scoping behavior, including cross-stage persistence.
We present a natural deduction system and a Kripke semantics for BML, and prove their soundness and completeness. We also provide a computational counterpart of BML as a typed lambda calculus for MSP. In addition to standard metatheoretic properties such as confluence and strong normalization, we develop a staged semantics for the calculus, showing that the type system supports the stage-by-stage execution model required for multi-stage programming.
arXiv:2602.10247v2 Announce Type: replace
Abstract: The Bayesian approach to inverse problems provides a practical way to solve ill-posed problems by augmenting the observation model with prior information. Due to its measure-theoretic underpinnings, the approach has raised theoretical interest, leading to a rather comprehensive description in infinite-dimensional function spaces. The goal of this article is to bridge the infinite-dimensional theory for linear inverse problems in distribution spaces and associated computational inverse problems without resorting to a discrete approximation of the forward model. We show that the discretization of the unknown of interest is not necessary for the numerical treatment of the problem, the only approximations required being numerical quadratures that are independent of any discrete representation of the unknown. To demonstrate the viability of the approach, an analysis of X-ray tomography inverse problem is given in the proposed framework, and an analysis of the connection between the proposed approach and a discretization-based one is also provided.
arXiv:2602.13780v3 Announce Type: replace
Abstract: Remote sensing (RS) change detection is essential for interpreting surface dynamics. Semantic change detection (SCD) further enables pixel-level understanding of multi-class transitions, yet remains sensitive to pseudo-changes induced by imaging conditions. Recent RS foundation models extract semantically consistent features across temporal and environmental variations, which is critical for mitigating pseudo-changes. However, existing SCD methods are often rigid and backbone-specific, lacking the flexibility to integrate diverse multi-scale features from emerging foundation models. To this end, we introduce a modular Cascaded Gated Decoder (CG-Decoder) that bridges various backbones and SCD tasks, processing multi-scale features in a coarse-to-fine manner while enabling adaptive change extraction. Building upon the RS foundation model PerA, we present PerASCD, a unified SCD framework. We further propose a Soft Semantic Consistency Loss (SSCLoss) to mitigate numerical instability in mixed-precision training. Extensive experiments on SECOND and LandsatSCD show that PerASCD achieves new state-of-the-art Sek scores (26.11% and 65.21%), surpassing the previous best by 0.61% and 4.95%, respectively. It also demonstrates exceptional data efficiency (outperforming the full-data baseline with 50% data), seamless cross-backbone generalization, and enhanced interpretability. Our approach maintains robust semantic consistency under radiometric variations, providing a reliable SCD solution. Code: https://github.com/SathShen/PerASCD.git.
arXiv:2607.18199v1 Announce Type: new
Abstract: Not all training samples contribute equally to large language model fine-tuning. Selecting informative training samples can reduce the computational cost while preserving downstream performance. Many existing data selection methods rely on indirect heuristics, such as data quality, diversity or reasoning trace length. However, the effectiveness of these fixed criteria is task-dependent and difficult to generalize across diverse downstream tasks. Perplexity-based data selection provides a simple and model-aware solution to estimate the sample difficulty, but existing approaches typically score the entire training sequence and ignore the difference in learning objectives of language modeling and reasoning tasks. In this paper, we propose PPL-Factory, a simple and interpretable data selection framework that combines task-aware perplexity-based scores and data budget-aware selection criteria. Experiments on GSM8K demonstrate that PPL-Factory outperforms other state-of-the-art data selection methods using only $1\%$ of the training set. With $10\%$ of the data, PPL-Factory exceeds full-data fine-tuning accuracy by 0.9 on GSM8K and 4.8 on MATH. Overall, our results demonstrate that task-aware and budget-aware perplexity-based selection provides an effective and applicable approach for efficient fine-tuning.
arXiv:2607.17907v1 Announce Type: cross
Abstract: The maximization of statistical complexity has long been associated with the emergence of probability distributions lying between perfect order and complete disorder. While previous studies have shown that complexity-maximizing distributions exhibit a two-level structure in finite discrete systems, an analogous unified treatment for both discrete and continuous probability spaces has remained unavailable. In this work, we develop a general variational framework for a generalized statistical complexity constructed from Shannon and Renyi entropies. We derive a common stationary equation governing both discrete probability masses and continuous probability densities and prove that every stationary solution necessarily possesses exactly two probability levels, establishing a universal core-halo structure. We further demonstrate that the optimization problem reduces to a single multiplicity parameter and prove that the global complexity maximum is attained by the smallest admissible core, corresponding to a single dominant state in the discrete case and an infinitesimal core in the continuous limit. These results provide a complete analytical characterization of the complexity-maximizing distributions and reveal a common mathematical structure underlying complexity optimization in both discrete and continuous settings. The framework establishes a unified foundation for generalized statistical complexity with potential applications in statistical mechanics, information theory, and the analysis of complex systems.
arXiv:2505.17964v2 Announce Type: replace
Abstract: Cycle count statistics are fundamental tools in statistics and engineering, with applications in motif counting, channel coding, and statistical inference of network and matrix data. However, how to compute high-order cycle count statistics efficiently is still an open problem. In this paper, we aim to derive Computationally Efficient Equivalent Forms (CEEF) for cycle count statistics of any given order, where we express each cycle count statistic equivalently as a linear combination of finitely many terms. Using the CEEF, we provide a much more efficient way to compute the cycle count statistics.
The CEEF problem has no known general solution and requires delicate combinatorial arguments together with extensive calculations. While this task is hard to accomplish by humans alone, it provides an ideal setting in which Artificial Intelligence (AI) can be useful. We solve the problem by combining several theorems we derive with powerful coding skills of modern AI systems. Our results leverage graph-theoretic arguments and yield new formulas for general cases that were previously unknown. We find that, although AI cannot solve the problem independently, it becomes highly effective when guided by humans through theorems we derive as well as a clear derivation strategy, step-by-step instructions, and carefully-written prompts.
We consider several statistical applications, including spiked matrix testing, estimation of weak spike eigenvalues, and pairwise network comparison. For each problem, we demonstrate that optimal statistical performance is achieved by using high-order cycle count statistics, and our CEEF formulas make their computation feasible on large-scale data sets.
arXiv:2505.24852v3 Announce Type: replace
Abstract: On-device learning at the edge enables low-latency, private personalization with improved long-term robustness and reduced maintenance costs. Yet, achieving scalable, low-power end-to-end on-chip learning, especially from real-world sequential data with a limited number of examples, is an open challenge. Indeed, accelerators supporting error backpropagation optimize for learning performance at the expense of inference efficiency, while simplified learning algorithms often fail to reach acceptable accuracy targets. In this work, we present Chameleon, leveraging three key contributions to solve these challenges. (i) A unified learning and inference architecture supports few-shot learning (FSL), continual learning (CL) and inference at only 0.5% area overhead to the inference logic. (ii) Long temporal dependencies are efficiently captured with temporal convolutional networks (TCNs), enabling the first demonstration of end-to-end on-chip FSL and CL on sequential data and inference on 16-kHz raw audio. (iii) A dual-mode, multiplier-free compute array allows either matching the power consumption of state-of-the-art inference-only keyword spotting (KWS) accelerators or enabling $4.3\times$ higher peak GOPS. Fabricated in 40-nm CMOS, Chameleon sets new accuracy records on Omniglot for end-to-end on-chip FSL (96.8%, 5-way 1-shot, 98.8%, 5-way 5-shot) and CL (82.2% final accuracy for learning 250 classes with 10 shots), while maintaining an inference accuracy of 93.3% on the 12-class Google Speech Commands dataset at an extreme-edge power budget of 3.1 $\mu$W.
arXiv:2602.13795v2 Announce Type: replace
Abstract: Large Language Models (LLMs) are accelerating the shift from an Internet of information to an Internet of Agents (IoA), where autonomous entities discover services, negotiate, execute tasks, and exchange value. Yet today's agents are still confined to platform silos and proprietary interfaces, lacking a common stack for interoperability, trust, and pay-per-use settlement. This article proposes \textit{Agent-OSI}, a functional interoperability architecture for a decentralized IoA, whose core contribution is agent-to-agent (A2A) communication and a Web-compatible, backend-agnostic settlement protocol built on HTTP 402 (Payment Required); identity, verifiable execution, and semantic orchestration are treated as boundary layers with interfaces to existing standards. We treat HTTP 402 as an application-layer challenge-response primitive -- analogous to HTTP 401 for authentication -- whose settlement backend (escrow contract, payment channel, or signed off-chain receipt) is a pluggable choice, instantiated via a blockchain escrow in our prototype. We implement a prototype and evaluate its communication and settlement performance. Results show that, for generative workloads, end-to-end latency is dominated by task execution rather than settlement confirmation, and that keeping negotiation and delivery off the settlement backend reduces per-session settlement cost by approximately 51\% relative to a more on-chain baseline.
arXiv:2602.13811v2 Announce Type: replace
Abstract: Physics-Informed Neural Networks present a novel approach in SciML that integrates physical laws in the form of partial differential equations directly into the NN through soft constraints in the loss function. This work studies the application of PINNs to solve a one dimensional coupled electro-elastodynamic system modeling linear piezoelectricity in stress-charge form, governed by elastodynamic and electrodynamic equations. Our simulation employs a feedforward architecture, mapping space-time coordinates to mechanical displacement and electric potential. Our PINN model achieved global relative L2 errors of 2.34 and 4.87 percent for displacement and electric potential respectively. The results validate PINNs as effective mesh free solvers for coupled time-dependent PDE systems, though challenges remain regarding error accumulation and stiffness in coupled eigenvalue systems.
arXiv:2602.14161v2 Announce Type: replace
Abstract: Detecting prompt injection, jailbreak attacks, and harmful requests is critical for deploying LLM-based agents safely, yet current evaluation practices in this literature overestimate generalization. We train activation-based classifiers (linear probes on LLM hidden states) on a benchmark of 18 datasets (prompt attacks plus benign sources) and propose Leave-One-Dataset-Out (LODO) evaluation, where the held-out dataset is never seen during training. Across four LLMs from three families (Llama-3.1-8B, Gemma-3-27B, Qwen-3.5-2B/4B), standard cross-validation reports a pooled AUC 8.0-16.5 points higher than LODO; per-dataset held-out-test-vs-LODO accuracy gaps span 1-25 points.
To understand the gap, we analyze the LODO stability of a linear probe's per-feature classifier coefficients, defining a retention metric for sparse-autoencoder (SAE) features that flags dataset-dependent shortcuts. 28-44% of top SAE features are shortcuts across models, a dataset-identity classifier reaches 96.6%, and the dataset-identifying and safety-relevant subspaces partially overlap. Standard domain-generalization fixes such as adversarial training, subspace projection, sample reweighting, and class balancing do not close the gap.
Finally, we show LODO-weighted SAE attributions filter dataset artifacts for more reliable per-prompt explanations. We release our framework at https://github.com/maxf-zn/prompt-mining so future prompt-attack classifiers can be evaluated under LODO alongside CV.
arXiv:2602.18528v4 Announce Type: replace
Abstract: Audio-visual continual test-time adaptation involves continually adapting a source audio-visual model at test-time, to unlabeled non-stationary domains, where either or both modalities can be distributionally shifted, which hampers online cross-modal learning and eventually leads to poor accuracy. While previous works have tackled this problem, we find that SOTA methods suffer from catastrophic forgetting where the model's performance drops well below even the source model due to continual parameter updates at test-time. In this work, we first show that adapting only the modality fusion layer to a target domain not only improves performance on that domain but can also enhance performance on subsequent domains. Based on this strong cross-task transferability of the fusion layer's parameters, we propose a method, $\texttt{AVReCAP}$, that improves test-time performance of the models without access to any source data. Our approach works by using a selective parameter retrieval mechanism that dynamically retrieves the best fusion layer parameters from a buffer using only a small batch of test data. These parameters are then integrated into the model, adapted to the current test distribution, and saved back for future use. Extensive experiments on benchmark datasets involving unimodal and bimodal corruptions show our proposed $\texttt{AVReCAP}$ significantly outperforms existing methods while minimizing catastrophic forgetting.
arXiv:2602.20575v2 Announce Type: replace
Abstract: Driving interaction data are important for training and evaluating autonomous drivingVision-Language-Action (VLA) models, but existing datasets contain limited denseinteraction samples and weak alignment between trajectories, visual inputs, and languageannotations. This work presents the Interactive Enhanced Driving Dataset (IEDD), alarge-scale interaction-oriented dataset constructed from five naturalistic trajectory datasets:Lyft Level 5, Waymo, nuPlan, INTERACTION, and SIND. IEDD contains 7.31 millionego-centric interaction segments, including 6.66 million multi agents cases, covering head-on,car-following, merging, and crossing interactions. Each segment is associated withtrajectory-derived interaction metrics describing interaction intensity and efficiency. Based onthese annotations, IEDD-VQA further provides trajectory-reconstructed BEV videos,structured interaction semantics, and multi-turn question-answer pairs. The dataset cansupport interaction mining, long-tail scenario analysis, VLA instruction tuning, andhierarchical evaluation of perception, behavior description, physical quantification, andcounterfactual reasoning.