Gerberas, such as the one seen here, are typically decoys for Russia's version of the Shahed attack drone. Danylo Antoniuk/Ukrinform/NurPhoto via Getty Images Ukraine is finding advanced antennas on Russia's low-priority foam and plywood Gerbera drones. A top drone advisor said it's a sign Russia has surged production of the anti-jamming component. That's just after Russia was facing a three-to-five-month wait for these antennas last year, he said. Ukrainian troops are finding advanced antennas with stronger anti-jamming features on Russian decoy drones, a sign that the Kremlin is overcoming a shortage of the key component. Serhii "Flash" Beskrestnov, a prominent Ukrainian drone analyst, posted photos to social media on Wednesday of a downed delta-wing Gerbera drone and a 12-element Controlled Reception Pattern Antenna. "Recently, 12-element Kometa antennas have begun to appear on the Gerberas," wrote Beskrestnov, who is the Ukrainian defense minister's advisor on battle drone tech. "For me, this is a sign that the plant that produces Kometa has increased its production capacity. A year ago, there were not enough Kometas even for KABs," he added, referring to Russian precision glide bombs. "There was a three-to-five-month queue at the plant." Antennas with more elements are generally more resistant to electronic warfare, one of Ukraine's main ways to disrupt or take down incoming drones. More basic versions of the Kometa, Russia's family of anti-jamming antennas, have four- or eight-element arrays. That a Russian Gerbera would be equipped with a 12-element antenna is significant. Moscow typically uses Gerberas as decoys to absorb Ukrainian air defense munitions hunting down the Geran-2, Russia's domestically produced analog of the Iranian Shahed-136 long-range attack drone. Often made of foam or plywood, the Gerbera is more expendable than the Geran-2, which has an estimated price of between $35,000 and $80,000 each. Gerberas cost about $10,000 each to make, Ukrainian officials estimated in 2024. Though Russia has, more recently, been seen equipping Gerberas with small warheads or using them as scouting tools, the propeller-driven drones are typically considered distractions or low-priority targets. Beskrestnov wrote that Russia has traditionally saved heavily on costs for the Gerbera, initially launching them without jamming protection at all. It later began fitting cheaper, Iranian four-element antennas to the drones, he added. Russia sometimes equips its attack drones with even more advanced, 16-element antennas — including some of Chinese make — but Beskrestnov said on Wednesday that these were still being reserved for "important products." Earlier this month, Kyiv said it used its homegrown Flamingo missiles to strike a production facility for anti-jamming antennas, the VNIIR-Progress complex in the Russian city of Cheboksary, about 600 miles from the Ukrainian border. The plant has regularly come under attack from long-range Ukrainian drones in the past year, with at least four separate reported strikes in 2024. The extent of any lasting damage to the facility, however, is unclear. Read the original article on Business Insider

The strike disrupted travel for hundreds of thousands of commuters between Long Island and New York City. LIRR The Long Island Rail Road is set to resume service after the MTA reached a deal with striking workers. The strike disrupted up to 300,000 daily commuters traveling between Long Island and New York City. The MTA says it will continue to offer alternative travel options, urging commuters to continue WFH on Tuesday. New York's Long Island Rail Road plans to resume service at noon on Tuesday after the Metropolitan Transportation Authority reached a deal with union leaders. "Tonight, the @MTA eached a fair deal with the five LIRR unions that delivers raises for workers while protecting riders and taxpayers," New York Gov. Kathy Hochul said on X. About 3,500 LIRR employees walked off the job shortly after midnight on Saturday following a breakdown in negotiations over wages and healthcare costs. The strike shut down North America's busiest commuter railroad, disrupting travel for about 300,000 daily riders traveling between Long Island and New York City The MTA said it will continue offering alternative travel options, including shuttle buses, as service ramps back up. Officials also urged commuters to continue working from home on Tuesday. The walkout marked the LIRR's first strike since 1994. The development came just days before the Memorial Day weekend, one of the busiest travel periods of the year, when many New Yorkers rely on the LIRR to reach the Hamptons — the affluent beach destination at the eastern end of Long Island, about 70 miles from Manhattan. Read the original article on Business Insider

The author said that she started making overpayments on her student loans once her oldest child graduated from college in 2022. Courtesy of Trisha Daab. As a single mom earning $36,000, my manager said grad school was the only way to earn more. When I graduated, I had about $69,000 in student loan debt for undergrad and grad school. Once my oldest graduated, I started paying more toward my loans and paid them off early. The year was 2006. Pluto had just been downgraded to a dwarf planet. I was obsessed with my Motorola Pebl flip phone and iPod. It was also the year I started paying down my student loan balance after completing my MBA. For 20 years, this student loan was a constant in my life, a staple of my monthly budget. It was older than my youngest child, and it was tied to a career I left in 2023. Finally paying it off in 2026 was a moment that I couldn't even fathom in 2006. I was told I need to go to grad school to get a promotion After earning my bachelor's degree, I landed an entry-level communications job at a large public company. I was young and determined — a single mom making about $30,000. The author celebrating her 23rd birthday, the year she started grad school, with her then 3-year-old. Courtesy of Trisha Daab. After a year, that salary increased to around $36,000, but I was eager to get promoted and make even more money. A manager gave me some advice: I would need a master's degree to have any hope of moving up. I listened and started grad school in 2003. The loans quickly added up Three years later, I had my MBA and a bill for over $69,000. This was the combined total of my $20,000 in unsubsidized undergrad and over $49,000 subsidized grad-school loans. The author, after her grad school graduation, with her husband at Navy Pier in Chicago. Courtesy of Trisha Daab. At the time, that was a terrifying amount. After all, it was more money than I had ever made in a year. But that manager was right. Getting my MBA was the key to getting promoted, which happened in 2006 when I landed a new job with a better title at a new company. My salary and bonuses greatly increased. For almost 20 years, I made that monthly payment The amount of my payment that applied to interest varied greatly over the 20 years it took to repay my debt. I had a standard consolidated 30-year loan. My first loan payment was in December of 2006, and the automatic payments I set up would continue until November of 2036. Every single month, $345.19 was deducted from my checking account. The first two payments went entirely toward interest, which was hard to stomach. Over the years that shifted to about 50/50 principal and interest, and then the interest got lower and lower. I eventually started paying more than the minimum amount When my oldest son graduated from college in 2022, and I was no longer giving him money for school, I was able to make extra payments and put a portion of my annual bonuses toward my loan repayment. I was determined not to be paying my student loans into my mid-50s. I paid off the loan in early 2026. Rather than the estimated $108,000, the total amount paid was $98,442.88. It worked out to be $69,928.46 in principal and $28,514.52 in interest. I didn't realize how freeing it would be to pay off these loans My life changed enormously in the 20 years I had been paying off my student loans. I had another child. I got promotions, changed companies, and helped my son and three stepchildren get through college, all while paying my own student loans. I made a career change in 2023 and walked away from marketing and Corporate America — the reason I had gone to grad school in the first place. But every monthly student loan payment was a reminder of the investment I made in a career I no longer wanted. Some days, I felt irresponsible, paying for this degree I wasn't really using anymore. And now, two major life changes — my youngest graduating from high school and my husband retiring — are giving me more time and headspace to do something in addition to being a travel writer, the job I left my marketing career for a few years ago. Yes, I'm considering adding some more lines to my résumé. I celebrated this milestone with my youngest over coffee A few days after making that final payment, I received an email from the lender congratulating me that my student loan was paid off. The author said it felt a bit surreal once her loans had been completely paid off. Courtesy of Trisha Daab. That was the moment it sank in. This thing that had been a constant in my life for over 20 years — through a second child, promotions, writing a book, even making a career change — is now done. Paying off my student loans freed me from feeling an obligation to a career path I no longer want to pursue. Finally, my thoughts about what I want to do next changed. Celebrating over coffee with my youngest, they shared what they think I should do next: teach. I might give that a try. Time will tell what this next chapter will bring, but I'm glad my student loans won't be following me. Read the original article on Business Insider

arXiv:2605.16650v1 Announce Type: new Abstract: Evaluating multi-turn dialogue systems remains challenging because response quality depends not only on the current prompt, but also on previously established entities, claims, and conversational commitments. Existing automatic evaluators, including LLM-as-a-judge frameworks and embedding-based metrics, largely rely on flat or turn-isolated representations, making them less effective at detecting long-range issues such as contradiction, topic drift, and entity inconsistency. To address this, we propose SKG-Eval, a quasi-deterministic and interpretable framework that models dialogue as an evolving Semantic Knowledge Graph (SKG) of entities, relations, and commitments across turns. The framework incrementally updates the graph through structured triple extraction and computes three complementary signals: (i) local relevance, measuring alignment with the current prompt and optional reference; (ii) historical consistency, evaluating how newly introduced information connects to prior conversational context using graph-based and embedding-driven signals; and (iii) logical coherence, assessed by a geometric contradiction engine that detects cross-turn conflicts without relying on NLI models or LLM judges. These signals are adaptively fused and aggregated into a length-invariant session score via recency-weighted trend analysis. Across multiple benchmarks, SKG-Eval achieves higher correlation with human judgments and substantially improves detection of long-range inconsistencies in extended conversations. In addition, the framework produces explicit contradiction certificates and deterministic scores for fixed inputs, enabling reproducible and auditable evaluation. Overall, our results suggest that structured externalized state tracking through semantic knowledge graphs provides a scalable alternative to implicit reasoning in LLM-based dialogue evaluators.

arXiv:2605.16317v1 Announce Type: new Abstract: Accurate, unified models for event cameras (ECs) remain elusive, hampering calibration and algorithm design. We develop a foundational probabilistic model for EC event detection, grounded in photon statistics, that unifies the description of static scene noise events and step response curves (S-curves) within a single analytical framework. Three formulations of the probability distributions are derived, spanning all intensity regimes: exact Poisson, saddle-point, and Gaussian. The model reveals the underlying connection between these otherwise disparate EC behaviors and clarifies the interpretation of S-curves, which we show is more nuanced than selecting a fixed probability threshold. Based on this model, we propose Noise2Params, a method for determining camera-specific values of the log-contrast threshold $B$, the lux-to-photon conversion factor $\alpha$, and the leakage term $\theta$ (found to be intensity dependent), via error minimization against observed noise-event distributions. Noise2Params requires only recordings of static, uniform scenes, offering an experimentally accessible alternative to approaches that demand specialized dynamic light sources. We further support the validity the model by training convolutional neural networks (CNNs) on synthetic noise images generated from our distributions and evaluating their ability to reconstruct static scenes from experimental data. We further demonstrate the utility of our model by showing that CNNs incorporating synthetic data outperform those trained solely on experimental data. Our framework provides a quantitative foundation for EC calibration, noise-aware algorithm design, and applications in photon-limited regimes.

arXiv:2605.16373v1 Announce Type: new Abstract: Early and accurate diagnosis and lesion localization of bone infections are crucial for clinical treatment. PET-CT integrates anatomical information from CT with metabolic information from PET, making it an important imaging modality for diagnosing bone infections. However, accurate lesion segmentation remains challenging due to indistinct lesion boundaries and inconsistencies in annotations generated by different experts or automated systems. In this work, we investigate multimodal segmentation of bone infections under annotation discrepancy. We develop a bimodal end-to-end segmentation framework that integrates PET metabolic signals and CT bone-window anatomy through an early-fusion multimodal representation.To mitigate performance inflation caused by inter-slice correlation in small datasets, this study discards traditional two-dimensional evaluation methods and implements a rigorous patient-level 3D volumetric evaluation and cross-validation. Furthermore, instead of forcing a singular consensus, we propose a decoupled dual-source learning framework where parallel models are trained on independent expert annotations driven by high-sensitivity and high-specificity clinical intents. Experimental results objectively report performance variations at the patient level (Mean + SD and Mean - SD), demonstrating the effectiveness of multimodal PET-CT fusion. The cross-evaluation matrix quantitatively reveals how models successfully internalize distinct expert diagnostic philosophies, providing a robust, diversity-preserving paradigm for clinical AI deployment in bone infection segmentation.

arXiv:2605.16381v1 Announce Type: new Abstract: Proactive streaming video understanding requires models to continuously process video streams and decide when to respond, rather than merely what to respond. This naturally introduces a decision-making problem under partial observations, where models must balance early prediction against sufficient evidence. However, existing benchmarks largely follow a "see-then-answer" paradigm, where responses are triggered only after explicit evidence appears, effectively reducing proactive reasoning to delayed perception. As a result, they fail to evaluate a model's ability to make timely and reliable decisions under incomplete observations. Moreover, training proactive models is inherently challenging due to the extreme imbalance between silence and response signals in streaming trajectories, as well as the need to jointly optimize response correctness and timing. To address these challenges, we introduce StreamPro-Bench, a new benchmark that evaluates streaming models from three complementary perspectives: Perception Understanding, Temporal Reasoning, and Proactive Agency, where the last measures a model's ability to make early yet reliable decisions under partial observations. We further propose StreamPro, a two-stage training framework for proactive learning. First, we introduce CB-Stream Loss to mitigate the severe supervision imbalance during supervised fine-tuning (SFT). Then, we apply Group Relative Policy Optimization (GRPO) with a multi-grained reward design that involves both turn-level and trajectory-level rewards. Experiments show that StreamPro significantly improves proactive performance. On StreamPro-Bench, it achieves 41.5, substantially outperforming the previous best (10.4), while also maintaining strong performance on real-time streaming benchmarks, achieving 78.9 on StreamingBench-RTVU.

arXiv:2605.16562v1 Announce Type: new Abstract: We report on the ongoing development of arXiv's HTML Papers offering, available on every new TeX/LaTeX submission since its initial release in 2023. The main highlights from 2025 and early 2026 are: (i) community-driven improvements to HTML fidelity and service health, with roughly half of 6,000 user reports resolved; (ii) corpus-scale conversion work aimed at 90% error-free HTML (currently 75%); (iii) initial MathML 4 Intent annotations for accessible speech output; (iv) an in-progress Rust port of LaTeXML, reducing compute costs and enabling faster previews on submission. The arXiv HTML Papers project remains experimental, but is gradually maturing as we better understand the needs of arXiv's readers and the technical opportunities presented by new standards and by advances in programming languages and AI.

arXiv:2605.16259v1 Announce Type: new Abstract: While real-time image generation using diffusion models has advanced rapidly on NVIDIA GPUs, systematic optimization research on non-CUDA platforms such as Apple Silicon remains extremely limited. In this study, we conducted comprehensive optimization experiments across 10 phases targeting the Apple M3 Ultra (60-core GPU, 512 GB unified memory) with the goal of achieving real-time camera img2img transformation. We explored a wide range of techniques including CoreML conversion, quantization, Token Merging, Neural Engine utilization, compact model exploration, frame interpolation, kNN search-based synthesis, pix2pix-turbo, optical flow frame skipping, and knowledge distillation, quantitatively evaluating the effectiveness of each approach. Ultimately, by combining CoreML conversion of the distillation-specialized model SDXS-512 with a 3-thread camera pipeline, we achieved real-time camera img2img transformation at 22.7 FPS at 512x512 resolution. The primary contribution of this work is the systematic demonstration that optimization insights established for CUDA are not necessarily effective on Apple Silicon's unified memory architecture. We reveal an optimization landscape fundamentally different from that of NVIDIA GPUs -- including the absence of speedup from quantization, the ineffectiveness of parallel inference, and the unsuitability of the Neural Engine for large-scale models -- and provide practical guidelines for diffusion model inference on Apple Silicon.

arXiv:2605.16319v1 Announce Type: new Abstract: Medium-horizon Alzheimer's disease progression prediction is difficult because future clinical scores can remain tied to baseline severity, while biomarker histories are irregular and incompletely observed. We develop an anchor-based analysis of 24-month Clinical Dementia Rating Sum of Boxes (CDR-SB) change using harmonized Alzheimer's Disease Neuroimaging Initiative (ADNI) tables. Each labeled sample is anchored at a mild cognitive impairment visit, uses only clinical and biomarker history observed at or before that anchor, and defines the response as CDR-SB at the future visit closest to 24 months within an 18--30 month window minus anchor CDR-SB. The analytic cohort contains 2,600 labeled anchors from 858 participants and 7,276 longitudinal rows. We propose a residual gap-aware transformer that combines a mixed-effects statistical reference with transformer-based residual learning from pre-anchor clinical and biomarker histories. The model uses participant-level random intercepts in the mixed-effects reference, observation-level triplet tokenization for irregular histories, and a learned nonnegative time-gap penalty inside self-attention. We compare the proposed model with a Bayesian-information-criterion-selected linear mixed-effects baseline, GRU-D, and STraTS under repeated participant-level train--test splits. Across five participant-level random seeds, the proposed model achieves the best mean test performance across all reported metrics, reducing MSE by 13.1% and increasing prediction--observation correlation by 26.4% relative to the mixed-effects baseline. It also improves over both GRU-D and STraTS in mean error and correlation. These results show that statistical anchoring and gap-aware residual learning provide a useful structure for medium-horizon Alzheimer's disease progression prediction.

arXiv:2605.16551v1 Announce Type: new Abstract: Evaluating LLM-based agents remains challenging because identifying meaningful failure cases often requires substantial human effort to design realistic test scenarios. Prior works primarily focus on automatically discovering agent failures induced by adversarial users, while overlooking queries with real user intents that also trigger agent failures. We introduce PQR, a framework that not only surfaces agent failures with respect to specific objectives (e.g., helpfulness, safety, etc.) but also resembles real users' intents. PQR operates through an iterative interaction between two complementary modules. The query refinement module performs rewrites to explore diverse query variations, while the prompt refinement module uses prior feedback to derive new objective-violating strategies and realism policies for refining prompts, which in turn generate failure-triggering yet realistic queries. We evaluate PQR on detecting an e-commerce QA agent's unhelpful responses. Our method uncovers 23% - 78% more unhelpful responses, and our generated queries are more diverse and realistic compared to previous methods.

arXiv:2605.16613v1 Announce Type: new Abstract: We introduce a novel approach to emotion modeling that shifts the focus from identification to evaluation, addressing the limitations of discrete classification in applied domains such as finance. By constructing a dataset of emotional intensity scores and fine-tuning open-weight generative language models to output continuous values from 0-100, we demonstrate a more expressive, generalizable framework for sentiment and emotion analysis. Our findings not only outperform classification baselines but also reveal surprising generalization capabilities and transfer effects to related constructs such as sentiment and arousal. This work contributes to the interdisciplinary recontextualization of NLP by introducing emotion intensity evaluation as an alternative to classification, arguing that this shift better aligns with the needs of domains--such as finance--where the degree of emotional content is central to interpretation and decision-making.

arXiv:2605.16654v1 Announce Type: new Abstract: Manner and result verbs encode different aspects of event structure and have been discussed in developmental work as a potentially informative distinction for studying early verb learning. However, this distinction remains difficult to measure at scale because large annotated resources for manner and result classification are not currently available. We present a computational approach for identifying manner and result verbs in sentence context. Using linguistically informed prompts, we generate sentence-level annotations with large language models over data drawn from MASC and InterCorp, extending coverage from previously annotated portions of VerbNet to 436 classes. We then train a RoBERTa-based classifier on these annotations and evaluate it on three held-out gold-standard datasets, including previously annotated items and a new expert-annotated set. Across these evaluations, the model shows promising performance, with average accuracy up to 89.6%. We present this work as a scalable measurement tool that can support future research on verb semantics in developmental and other language datasets, while noting that further validation is needed for borderline cases, mixed manner/result verbs, and downstream developmental applications.

arXiv:2605.16371v1 Announce Type: new Abstract: Large Multimodal Models (LMMs) often struggle with geometric reasoning due to visual hallucinations and a lack of mathematically precise Chain-of-Thought (CoT) data. To address this, we propose the GeoSym Engine, an automated and scalable neuro-symbolic framework. By leveraging a type-conditional grammar and an analytic SymGT Solver, it derives exact symbolic ground truths and seamlessly integrates with a robust rendering pipeline to produce high-precision geometric diagrams. Using this engine, we construct GeoSym127K, a difficulty-stratified dataset featuring 51K high-resolution images, 127K questions with symbolic ground truths, and 55K answer-verified CoT QA pairs. We also introduce GeoSym-Bench, an expert-curated suite of 511 complex samples for rigorous evaluation. Through extensive supervised fine-tuning (SFT), we demonstrate that GeoSym drives concentrated improvements specifically on diagram-dependent and multi-step geometry tasks. Our Qwen3-VL-8B model gains an absolute +22.21% on the MathVerse Vision-Only subset and reaches 61.52% (+6.19% improvement) on WeMath, mitigating long-horizon logic fragmentation and outperforming advanced closed-source models like Doubao-1.8. Furthermore, applying Reinforcement Learning with Verifiable Rewards (RLVR) via GRPO reveals that initializing from structural SFT checkpoints substantially elevates the performance ceiling over zero-shot RL. Driven by deterministic exact-match signals, this showcases the robust scaling potential of our verifiable reasoning synthesis. Datasets and code are available at https://huggingface.co/datasets/Tomie0506/GeoSym127K and https://github.com/Tomie56/GeoSym127K.

arXiv:2605.16758v1 Announce Type: new Abstract: Large Language Models (LLMs) remain substantially less data-efficient than humans. Pre-pretraining (PPT) on synthetic languages has been proposed to close this gap, with prior work emphasizing highly expressive formal languages such as $k$-Shuffle Dyck. Inspired by the Language Acquisition Device (LAD) hypothesis, which posits that innate constraints preemptively restrict the learner's hypothesis space to natural-language-like structure, we propose LAD-inspired PPT: pre-pretraining on MP-STRUCT, a formal language whose strings encode hierarchical composition, feature-based dependencies, and long-distance displacement via MERGE, AGREE, and MOVE. A brief 500-step PPT with MP-STRUCT matches strong formal-language baselines in token efficiency while additionally imparting a human-like resistance to structurally implausible languages (e.g., REVERSE). Analyzing simplified variants, we find that MP-STRUCT CORE outperforms $k$-Shuffle Dyck despite not being definable in C-RASP (a formal bound on transformer expressivity), challenging the prior hypothesis that effective PPT languages must be both hierarchically expressive and circuit-theoretically learnable. We show that functional landmarks, which reduce dependency resolution ambiguity, are a key driver, suggesting that effective PPT design depends not only on expressivity but also on the accessibility of dependency resolution.

arXiv:2605.16679v1 Announce Type: new Abstract: End-to-end automation of realistic healthcare operations stresses three capabilities underrepresented in current benchmarks: policy density, decisions must be grounded in a large library of medical, insurance, and operational rules; Multi-role composition: a single task requires the agent to play multiple roles with handoffs; and multilateral interaction: intermediate workflow steps are multi-turn dialogs, such as peer-to-peer review and patient outreach. We introduce $\chi$-Bench, a benchmark of long-horizon healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. Each task hands the agent a clinical case in a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, which it must drive to a terminal status through tool calls and writing the role's artifacts, guided by a 1,290+ document managed-care operations handbook skill. Across 30 agent harness/models configurations, the best agent resolves only 28.0% of tasks, no agent clears 20% on strict pass^3, and executing all tasks in a single session slumps the performance to 3.8%. These results raise the hypothesis that similar gaps are likely to surface in other policy-dense, role-composed, irreversible enterprise domains.

arXiv:2605.16353v1 Announce Type: new Abstract: Continual Visual Instruction Tuning (CVIT) enables Multimodal Large Language Models to incrementally acquire new abilities. However, existing CVIT methods operate under a restrictive task-incremental setting, where each training phase corresponds to a single, predefined task. This does not reflect real-world conditions, where data arrives as a continuous stream of interleaved and dynamically evolving tasks. To bridge this gap, we introduce Streaming CVIT (StrCVIT), a more general and realistic setting where models learn from a stream of data chunks containing a dynamic mixture of tasks. In StrCVIT, a model must simultaneously acquire new abilities, reinforce recurring abilities, and mitigate forgetting. Existing CVIT methods fail here as they cannot reliably distinguish or adapt to the heterogeneous task samples within each chunk. We therefore propose StrLoRA, a regularized two-stage expert routing framework. StrLoRA first performs task-aware expert selection using the textual instruction to activate a sparse subset of relevant experts, reducing cross-task interference. It then applies token-wise expert weighting within this subset, where contribution weights are computed via cross-modal attention between local visual tokens and the global instruction representation. To maintain stability across the non-stationary stream, a routing-stability regularization aligns current routing distributions with a historical exponential moving average reference. Extensive experiments on a newly developed StrCVIT benchmark show that StrLoRA substantially outperforms existing methods, effectively enhancing model's abilities from continuously evolving data streams.

arXiv:2605.16359v1 Announce Type: new Abstract: Vision-language models improve perception by feeding increasingly long visual token sequences into language backbones, but the resulting inference cost raises a basic scaling question: as multimodal models grow, how many visual tokens are actually needed, and how should they be allocated under a fixed visual token budget? Existing training-free pruning methods typically answer this with one-shot proxies such as decoder attention, visual similarity, or conditional diversity. We argue that visual token pruning is better viewed as task-conditioned evidence search, especially under aggressive compression and across model scales. We propose F^3A, a training-free router for visual token pruning that operates before the language model consumes image tokens. F^3A builds lightweight question-conditioned cues, matches them to visual-grid tokens through frozen sparse sensing heads, and allocates a fixed vision token budget via coarse evidence localization, local refinement, coverage-preserving competition, and recovery of under-covered regions. It requires no model training, no extra LLM forward pass and preserves the original multimodal prompting and decoding pipeline.

arXiv:2605.16372v1 Announce Type: new Abstract: Steering or intervening on model representations at inference time to correct predictions is essential for AI interpretability and safety, yet existing evaluation protocols are limited to ambiguous language modeling tasks. To address this gap, we introduce SwordBench, a benchmark for steering image representations of vision models across multiple backbones and concept removal tasks. Beyond a unified benchmarking suite, we propose new evaluation notions that uncover the second-order effects of orthogonalization among concept activation vectors for pragmatic steering. Specifically, cross-concept robustness measures the stability of concept detection performance across inputs orthogonalized against alternative concepts, and collateral damage quantifies whether steering inadvertently affects model performance on a downstream task for inputs lacking the bias. We find that although a linear support vector machine exhibits superior separability and orthogonality, it fails to achieve zero collateral damage, often trailing sparse autoencoders. In simpler regimes, both standard baselines and optimization-based methods fail to achieve perfect steering. The source code will be made available soon on GitHub.

arXiv:2605.16508v1 Announce Type: new Abstract: As agent systems scale, skills accumulate into large reusable libraries, yet their scaling laws remain poorly understood. Across 15 frontier LLMs, 1,141 real-world skills, and over 3M routing or execution decisions, we identify two coupled laws. Routing law: single-step routing accuracy decays logarithmically with library size ($R^2{>}0.97$ for all models), with errors progressing from local skill competition to cross-family drift and capture by overly general "black-hole skills". Execution law: before state realization, joint routing is approximately multiplicative, whereas correct execution can improve difficult downstream decisions by about $4{\times}$. A single parameter, the routing logarithmic decay slope $b$, couples the two laws: routing-side fits predict execution-side rescue across models, showing that the same library property controls both pre-execution collapse and downstream recoverability. The laws are actionable: law-guided optimization raises held-out routing accuracy from 71.3% to 91.7%, reduces hijack from 22.4% to 4.1%, and transfers directionally to downstream ClawBench and ClawMark execution settings, improving mean pass rate from 49.3% to 61.6% on ClawBench and from 28.4% to 34.5% on ClawMark. These results show that agent performance depends not only on model capability, but also on the structure, granularity, and exposure policy of the skill library.