Discussion | Link

Satya Nadella isn't a big fan of tokenmaxxing for the sake of tokenmaxxing. Sven Hoppe/picture alliance via Getty Images Microsoft CEO Satya Nadella said "a lot" of tokenmaxxing is happening at Microsoft. He said workers should use the right AI model for the job. Nadella said he vibe-coded a tool that updates code from workplace chats. Satya Nadella has a message for Microsoft's AI-pilled employees: Not every problem needs the most powerful AI model. At a live taping of The New York Times' "Hard Fork" podcast, Microsoft's CEO was asked how much tokenmaxxing is happening inside the company. "A lot," Nadella said, before cohost Casey Newton could finish the question. "I'm a tokenmaxxer too, it's addictive," Nadella said. "But you have to step back when the novelty wears off to say, 'What is it that I'm trying to create?'" Silicon Valley executives have spent the past year pushing workers to use AI as much as possible, sometimes through internal leaderboards that track tokens, the units of data processed by AI systems. Now that the bills are due, companies are putting AI use on a diet. Nadella did not say Microsoft is limiting employees' AI use, but he said workers should use the right model for the job. "Don't use frontier models for non-frontier problems," Nadella said, pointing to Microsoft Copilot's auto mode, which is designed to match tasks with the model most appropriate for it. "Let's kind of match these things such that you get the outputs, you get the economics — it can't be a race to doing things that just don't add value." Nadella added that he recently vibe-coded a tool that keeps a software project up to date by following related workplace conversations. If employees discuss a change connected to the project, the AI can create a plan, make the update, and keep the code working without Nadella needing to be in the meeting or thread. Nadella has been working to remake Microsoft for the AI era and turn the 220,000-person company into one that can compete with smaller, faster rivals. In October, he appointed a new CEO of Microsoft's commercial business, a move that freed him up to spend more time on technical work. In November, he tapped a new AI advisor to help rethink the company's business model for the AI era. During the live event, cohost Kevin Roose presented Nadella with what he called a "piece of rare merchandise": a T-shirt that read "Microsoft Advanced AI Research." Roose said he had acquired it from an OpenAI employee who had it made in 2023, when Sam Altman was briefly ousted, and Microsoft was preparing to create a new AI lab for OpenAI employees. Altman was reinstated days later, and the lab was never created. Nadella, laughing, accepted the T-shirt. Read the original article on Business Insider

The decision comes as India emerges as the world’s largest GCC market.

Game 4 of the NBA Finals between the New York Knicks and San Antonio Spurs took place at Madison Square Garden in New York City. Al Bello/Getty Images The Knicks and Spurs returned to Madison Square Garden in New York City for Game 4 of the NBA Finals. Taylor Swift, Kylie Jenner, and Adam Sandler made their first appearances on Wednesday night. Here are the celebrities, sports personalities, and business executives we spotted at the NBA Finals. The NBA Finals remained in New York for Game 4 on Wednesday night, bringing another star-studded crowd to Madison Square Garden. The Knicks edged the Spurs 107-106 on a go-ahead basket in the final five seconds, taking a 3-1 lead in the series. Game 5 is scheduled for Saturday. Here are the politicians, business executives, celebrities, and sports personalities we spotted at the NBA Finals so far. Taylor Swift and Mariska Hargitay Alana Haim, Este Haim, Taylor Swift, and Mariska Hargitay were spotted courtside at Game 4 of the 2026 NBA Finals. Al Bello/Getty Images Taylor Swift was spotted courtside at Game 4 with fellow musicians Alana and Este Haim of the band Haim, as well as "Law and Order: SVU" star Mariska Hargitay. The quartet arrived in matching Knicks-inspired shirts: "Stevie Knicks" for Swift and Hargitay, "Knickelback" for Alana Haim, and "Knickole Kidman" for Este Haim. Hargitay, a longtime Knicks supporter who also attended Game 3, has a big fan on the home team. Knicks captain Jalen Brunson has previously spoken about being a fan of "Law & Order: SVU." "I've always loved that show," Brunson said during a May 2025 episode of "The Roommates Show" podcast, which he hosts with his teammate Josh Hart. Michael J. Fox and Tracy Pollan Michael J. Fox and his wife, Tracy Pollan, were at Game 4 of the NBA Finals. Al Bello/Getty Images Michael J. Fox, one of the Knicks' most devoted celebrity fans, attended Game 4 with his wife, Tracy Pollan. Speaking to People in an interview published on Monday, Fox said he has a game-day ritual he never skips. "I always get a fist bump from Spike [Lee] on the way to my seat and a big sloppy hug from Tracy Morgan," Fox said. Adam Sandler Adam Sandler was at Game 4 of the NBA Finals. Al Bello/Getty Images Adam Sandler, a longtime Knicks fan, was spotted at Madison Square Garden in an orange Knicks hoodie. Chris Rock Chris Rock is at MSG for Knicks-Spurs Game 4! pic.twitter.com/BbQNpknz0L — SNY Knicks (@sny_knicks) June 11, 2026 Chris Rock was present at Wednesday's game, wearing Knicks colors. Jerry Seinfeld Jerry Seinfeld attended Game 4 of the NBA Finals with his wife, Jessica Seinfeld. Dustin Satloff/Getty Images Jerry Seinfeld and his wife, Jessica, were spotted courtside at Game 4. Timothée Chalamet and Kylie Jenner Timothée Chalamet and Kylie Jenner pulled up to watch the Knicks in matching Chrome Hearts 🔥 pic.twitter.com/aRP9EXqjDi — SportsCenter (@SportsCenter) June 11, 2026 Knicks supporter Timothée Chalamet, who has attended every game of the Finals so far, was back at Madison Square Garden for Game 4. This time, he was joined by Kylie Jenner. The couple wore matching Chrome Hearts outfits inspired by the Knicks' team colors, as seen in a video uploaded on X by ESPN SportsCenter. David Solomon Goldman Sachs CEO David M. Solomon was at Game 4 of the NBA Finals. Al Bello/Getty Images Goldman Sachs CEO David Solomon joined the star-studded crowd at Madison Square Garden for Game 4 of the NBA Finals. Jimmy Fallon Jimmy Fallon hyping up the Knicks crew at Game 4! 👏 Presented by Tri State Cadillac. #sponsored pic.twitter.com/GgpcSQ3BDd — New York Post Sports (@nypostsports) June 11, 2026 Jimmy Fallon was seen making his way through the arena wearing a "New York" hoodie, per a video uploaded on X by the New York Post. Wu-Tang Clan Wu-Tang Clan performed at halftime of Game 4 of the 2026 NBA Finals. Al Bello/Getty Images Wu-Tang Clan took over Madison Square Garden at halftime of Game 4, delivering an energetic set that included hits such like "Wu-Tang Clan Ain't Nuthing ta F' Wit" and "Method Man." Ben Stiller and Christine Taylor Ben Stiller attended Game 4 of the NBA Finals with his wife, Christine Taylor. Al Bello/Getty Images Longtime Knicks fan Ben Stiller has attended every game of the NBA Finals so far, traveling to San Antonio for Games 1 and 2 before cheering on the team at Madison Square Garden for Games 3 and 4. Stiller attended Wednesday's game with his wife, Christine Taylor, and the pair were seated courtside alongside Swift and the Haim sisters. Spike Lee Spike Lee was at Game 4 of the NBA Finals. Al Bello/Getty Images Filmmaker Spike Lee hasn't missed a Finals game, making the trip to San Antonio before returning to Madison Square Garden for Games 3 and 4. In an interview with ESPN on Monday, Lee said he had read that his seat for Game 3 could have fetched a hefty sum on the secondary market. "It went down, but it was half a million," Lee said. In 2024, Lee was recognized as a SuperFan by the Naismith Basketball Hall of Fame. He has held New York Knicks season tickets since 1985, per ESPN. The filmmaker was also photographed posing with former Knicks guard Jeremy Lin, who attended the game. Tracy Morgan Tracy Morgan at Game 4 of the NBA Finals. Al Bello/Getty Images Like Stiller, Chalamet, and Lee, Tracy Morgan has been a constant presence throughout the NBA Finals. He was in San Antonio for Games 1 and 2, and was back in New York for Games 3 and 4. Morgan has long been one of the Knicks' most outspoken celebrity supporters. In a May video with Complex Sneakers, Morgan said he bought the team two pairs of Nike Kobe 5 "Bruce Lee" sneakers each in 2024. "Just to let them know I'm part of the fanbase and we appreciate what they do," Morgan said in the video. Morgan recently lent his voice to special New York City subway announcements encouraging fans to take the train to Knicks watch parties and home games at Madison Square Garden. Larry David and David Zaslav David Zaslav and Larry David were at Game 4 of the NBA Finals. Dustin Satloff/Getty Images Warner Bros. Discovery CEO and president David Zaslav was courtside with Larry David at Game 4. Both of them were also at Game 3 earlier this week. In July 2024, The Athletic reported that the NBA had reached a new 11-year, $76 billion media rights agreement with Disney, NBC, and Amazon — a deal that ultimately brought Warner Bros. Discovery's 35-year partnership with the league to an end after the 2024-25 season. At a Morgan Stanley conference in March 2025, Zaslav said Warner Bros. Discovery was better off investing in its own franchises rather than sports rights, which he described as a "rental business." Celebrity sightings from Game 3 President Donald Trump and his granddaughter, Kai Trump, at Game 3 of the NBA finals at Madison Square Garden. SAUL LOEB / AFP via Getty Images For Game 3, on Monday night, President Donald Trump was spotted at Madison Square Garden alongside his granddaughter, Kai Trump. During a pregame interview on the "NBA Tip-off," NBA commissioner Adam Silver said that Trump was a guest of Knicks' owner James Dolan. This marks the first NBA Finals game attended by a sitting US president, per ESPN. He wasn't the only politician in attendance: New York City Mayor Zohran Mamdani was also spotted fist-bumping fans as he entered the arena, while Michael Bloomberg briefly found himself in the middle of the action when the Knicks' point guard Jose Alvarado tumbled into his courtside seat during a play. Knicks fans Stiller, Lee, Morgan, and Chalamet were once again in the crowd after making appearances at Games 1 and 2. Stiller was with his wife, Christine Taylor, while Morgan brought along his former "30 Rock" costar Tina Fey. Other New York icons in attendance included Jay-Z and Derek Jeter, along with former Knicks guard Jeremy Lin. Hargitay and her former "Law and Order: SVU" costar Christopher Meloni were also at the game on Monday night. New England Patriots owner Robert Kraft was also courtside with Zaslav and David. Celebrity sightings from Game 2 Ben Stiller also attended Game 2 of the NBA Finals. Gregory Shamus/Getty Images Game 2 of the NBA Finals took place on June 5, and there were several familiar faces back in the crowd, including Stiller, Lee, Morgan, and Chalamet — all of whom were also at Game 1. Twelve-time NBA All-Star Chris Paul was among the attendees at Game 2. He played for the Spurs during the 2024-25 season and brought his son, Chris Paul II, to the game. Celebrity sightings from Game 1 Disney's Bob Iger, ESPN chairman Jimmy Pitaro, and other media executives were spotted at Game 1 of the NBA Finals in San Antonio, Texas. Screengrab by Steven Tweedie. The NBA Finals began on June 3 at Frost Bank Center in Texas. The Knicks, making their first Finals appearance in more than two decades. Several notable Knicks supporters were in attendance, including Stiller, Lee, Morgan, and Chalamet. They were joined by former New York Rangers goaltender Henrik Lundqvist and NBA legend Shaquille O'Neal, one of the hosts of the NBA Finals pregame show "NBA Tip-Off" and the postgame show "Inside the NBA." Disney board member Bob Iger, Disney CEO Josh D'Amaro, and ESPN chairman Jimmy Pitaro were also present at the game, together with Silver. During the broadcast, the commentators for the night joked that the men made up a "murderers' row of executives." The NBA Finals air on ABC and are produced by ESPN, two Disney-owned networks. Read the original article on Business Insider

arXiv:2606.11206v1 Announce Type: new Abstract: Supervised Fine-Tuning (SFT) is the predominant paradigm for aligning large language models (LLMs), yet it suffers from optimization instability and limited generalization. Recent work attributes this issue to pathological gradient scaling and proposes Dynamic Fine-Tuning (DFT) to correct it at the token level. However, DFT assumes all demonstrations are equally suitable learning targets, an assumption violated by the strong heterogeneity of large-scale instruction data, where demonstration-policy mismatch induces high-variance updates at the sample level. We introduce Compatibility-Aware Dynamic Fine-Tuning (CADFT), a principled extension of DFT that controls sample-level optimization variance. CADFT derives a dynamic, policy-dependent compatibility signal from model likelihoods to modulate supervised updates, suppressing high-variance gradients from incompatible demonstrations. We further propose a delayed, low-frequency compatibility-guided rewriting strategy to transform persistently incompatible demonstrations into learnable targets. We show that CADFT can be interpreted as a variance-controlled estimator that generalizes token-level stabilization in DFT to the sample level. Extensive experiments demonstrate improved stability, generalization, and cold-start reinforcement learning initialization, while remaining fully supervised and independent of explicit reward modeling.

arXiv:2606.11204v1 Announce Type: new Abstract: Accurate extraction of structured information from Safety Data Sheets (SDS) remains challenging in industrial safety due to heterogeneous document formats and the limitations of traditional rule-based methods. This study benchmarks state-of-the-art Large Language Models (LLMs) for automated SDS data extraction, comparing text-based and multimodal processing pipelines. We systematically evaluate four models: Gemini 1.5 Pro, GPT-4o, Claude 3.7 Sonnet, and Llama 3.1-70B, across three prompting strategies: zero-shot, few-shot, and chain-of-thought. The evaluation framework assessed accuracy, latency, and cost across more than 50,000 extracted data fields. Results show that text-based extraction consistently outperforms multimodal processing across all metrics. Gemini 1.5 Pro combined with a Chain-of-Thought prompt achieved the highest accuracy (84%), outperforming GPT-4o (81%) and Claude 3.7 Sonnet (79%). However, no model surpassed the 90% accuracy threshold commonly required for reliable real-world deployment. These findings indicate that general-purpose LLMs are not yet robust enough for unsupervised industrial use, though performance suggests strong potential with task-specific fine-tuning. Future research should focus on domain-adapted training, model calibration, and the integration of Human-in-the-Loop verification to ensure safety-critical reliability.

arXiv:2606.11196v1 Announce Type: new Abstract: Decentralized LLM inference networks need lightweight, reference-free quality evaluation for Proof of Quality (PoQ). We present PoQ-Judge, a framework that trains dedicated judge models to score query-output pairs without ground-truth references. We study three architectures across the quality-cost tradeoff: a TextCNN judge, a MiniLM cross-encoder, and a DeBERTa judge. Using two-stage training on UltraFeedback plus GPT-labeled in-domain data, the best model reaches 0.747 Pearson correlation with the ground-truth proxy on a held-out test set, outperforming reference-based evaluators from prior work. As a reference-free component in composite scoring, it achieves 0.645 Pearson correlation, matching the best single reference-based evaluator while removing the need for reference answers. We also show that online calibration identifies semantic quality as the dominant dimension and that cascade evaluation reduces cost by 72.7 percent with only modest quality loss. Results are much stronger on QA than summarization, pointing to proxy quality as the main remaining limitation.

arXiv:2606.11199v1 Announce Type: new Abstract: We present NightFeats, a structured multi-agent retrieval-augmented generation (RAG) system submitted to the MMU-RAGent competition at NeurIPS 2025, where it was awarded Best Dynamic Evaluation in the text-to-text track. Rather than targeting benchmark maximization, this work proposes a principled pipeline that decomposes knowledge synthesis into three coordinated phases: retrieval, curation, and composition, each governed by explicit intermediate representations and handoff contracts. Inspired by Agentic Context Engineering (ACE), the system introduces temporal-semantic reranking, bounded contradiction reconciliation, and citation-preserving composition as core architectural primitives. Competition results show that NightFeats surpasses proprietary baselines including Claude-SonnetV2 and Nova-Pro on LLM-as-a-Judge and Human Likert evaluations, confirming that architectural transparency and verifiable evidence grounding are better aligned with human preferences than systems optimizing narrowly for automatic similarity metrics.

arXiv:2606.11289v1 Announce Type: new Abstract: Diffusion models have consistently driven progress in text-to-image generation. However, it is challenging to attribute recent progress to specific modeling and data choices: state-of-the-art open-weight models provide limited ablations, and do not disclose their training data and full training details. The research community needs fully open (weights, data, and code) models as a foundation for further research; yet existing fully open models still fall significantly short of leading models in performance. In this project, we conduct a systematic investigation of the modeling and data design choices in text-to-image diffusion training and inference with 300+ controlled experiments totaling 700K+ TPU v6e hours. Our experiments highlight several empirical findings (e.g., equal weighting is a strong default for mixing curated datasets) and simple design decisions (e.g., larger text encoder adapters improve performance with minimal added parameters) for training strong models. Guided by these insights, we train i1, a 3B-parameter text-to-image diffusion model using only publicly available datasets. i1 is competitive with leading models on five representative benchmarks (GenEval, DPG, PRISM, CVTG-2K, and LongText), and outperforms the best existing fully open model by 29.5 absolute percentage points on average. We provide the i1 checkpoints, training and inference code, and the data processing pipeline. Together, our findings and the i1 recipe establish a practical foundation for future open research in text-to-image diffusion models. Our code is available at https://github.com/zlab-princeton/i1.

arXiv:2606.11269v1 Announce Type: new Abstract: Personality assessment aims to infer stable personality traits from dynamic behaviors across language, voice, and facial cues. Since different personality dimensions are revealed through distinct behavioral perspectives, modeling trait-specific evidence is challenging. However, most existing approaches adopt a uniform multimodal fusion strategy across all dimensions, assuming identical modality contributions. This overlooks trait-specific modality preferences and introduces cross-modal interference. To address this issue, we propose a novel personality assessment framework called Traits Run Deeper, which consists of three components. Specifically, the Multimodal Foundation Representation (MFR) module constructs personality-oriented multimodal inputs and leverages psychology-informed semantic templates as anchors, enabling foundation models to capture trait-relevant information. Building upon MFR, the Trait-Specific Modality Fusion (TSMF) module acts as an asymmetric fusion mechanism, allowing each dimension to selectively exploit different modality pathways from modality-specific modeling to complementary fusion. Thus, TSMF captures heterogeneous modality preferences while reducing cross-modal contamination. Furthermore, the Distribution-Calibrated Personality Regression (DCPR) module mitigates label imbalance and central tendency bias through target distribution calibration, improving robustness and stability. Experimental results on the AVI Challenge 2026 validation set demonstrate the effectiveness of the proposed framework, reducing mean squared error (MSE) by approximately 25% compared with the baseline. Consistent improvements are observed on the official test set, where our method achieves the best performance and ranks first in the Personality Assessment Track. The source code will be made available at https://github.com/MSA-LMC/AVI2026.

arXiv:2606.11200v1 Announce Type: new Abstract: Generative AI has enabled the creation of photorealistic images and videos that are increasingly disseminated on social media, often used for spam, misinformation, manipulation, and fraud. Existing AI-generated content (AIGC) detection methods face challenges including poor generalization to new generation models, reliance on single modalities, and lack of interpretable explanations. We present our pipeline that mitigates these issues by continuously curating diverse multi-modal social media data and training a compact vision-language model for detection and explanation. Our model achieves state-of-the-art detection performance on public benchmarks and demonstrates robust detection and explanation capabilities on internal social media datasets across multiple platforms. We deployed our model for post recommendation on social media platforms and observed positive downstream impacts on user engagement, demonstrating that it is feasible to perform effective AIGC detection in dynamic, real-world social media environments.

arXiv:2606.11314v1 Announce Type: new Abstract: We introduce TRON, a rendering framework that combines 3D Gaussian ray tracing with neural rendering to enable realistic and controllable rendering of real-world 3D scenes under novel lighting, dynamic object motion, object insertion, and material editing. Prior approaches that rely solely on physically based rendering (PBR) of Gaussian representations struggle to achieve realistic relighting due to imperfections in reconstructed geometry, material estimates, and light transport estimation. At the same time, neural rendering methods often lack an explicit scene representation, limiting their ability to support interactive editing with fine-grained manipulation. TRON bridges these two paradigms. We use intrinsic decomposition priors from a learned inverse rendering model to regularize the material properties of a Gaussian field, and repurpose a ray tracer to provide radiometric guidance rather than final pixels. By treating this output as a structured 3D scaffold, we empower a lightweight neural renderer to bridge the domain gap between shading-model constrained estimates and photorealistic output. Our key insight is that the combination of explicit 3D knowledge with robust material priors provides speed and controllability, while neural rendering enables the synthesis of photorealistic images. To support real-world scenarios, we train our neural renderer with a multi-stage strategy consisting of large-scale pretraining and targeted fine-tuning on a newly constructed dataset of 2.1M rendered synthetic and real-world frames from 3D reconstructions. TRON outperforms Gaussian-based relighting methods in realism, and prior neural renderers in editability and speed. To the best of our knowledge, TRON is the first method to enable practical interactive applications in captured 3D environments, offering realistic appearance under dynamic geometric, lighting and material conditions.

arXiv:2606.11198v1 Announce Type: new Abstract: Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution. We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3x more attention per token than semantically equivalent natural-language text ($\hat{o}$(KG) $\approx$ 0.70 vs. $\hat{o}$(neutral) $\approx$ 0.25), compressing demonstration attention by up to 42% -- regardless of whether the triples are relevant or noise. We develop a formal framework decomposing attention scores into semantic and structural components (Eq. 2), derive a compression bound (Proposition 1) connecting token-level format bias to demonstration attention loss, and show that the structural term governs how much attention is diverted while the semantic term governs whether this helps or hurts. This decoupling reveals two orthogonal axes for improving retrieval-augmented ICL: optimising retrieval quality (semantic axis) and reducing format-driven attention capture (structural axis). Empirically, across two model families (Mistral-7B, LLaMA-3-8B) and three QA benchmarks, we observe that source-task alignment dominates: task-matched BM25 retrieval achieves 58-62% on HotpotQA vs. ConceptNet's 25-27%, a >30 pp gap that dwarfs all gating strategies ($\leq$2 pp). We derive five structure-aware mitigation strategies from the framework, ranging from zero-cost prompt modifications to training-time regularisation; format flattening (S3) is validated by both accuracy and attention-level evidence from a verbalized-triple control, while structural dispersal (S1) yields mixed results that illuminate the challenges of format-level intervention.

arXiv:2606.11203v1 Announce Type: new Abstract: Structured sequence generation often requires a model to satisfy several input-derived constraints in a single output. Standard decoding methods may assign high probability to fluent continuations while placing low mass on continuations that realize all required anchors jointly. We study this regime as a rare-event sequential inference problem. LatticeBridge combines a compact prefix language model, instance-compiled surface automata, and a twisted sequential Monte Carlo (SMC) decoder with resampling, multilevel splitting, and a source-support proposal term derived from instance-provided phrases. The constraint representation is compiled from each input instance and does not rely on manually curated lexical classes. On 2,610 attainable validation tasks spanning CommonGen, E2E NLG, and WikiBio, the particle decoder improves exact anchor satisfaction and mean anchor coverage over greedy, beam-filtered, and best-of-k ancestral baselines under a shared proposal model. Since exact anchor satisfaction alone does not rule out unsupported attribute substitutions, the evaluation reports required-anchor coverage, source coverage, source-intrusion diagnostics, overlap, runtime, and particle statistics jointly. The benchmark characterizes the faithfulness-overlap-latency frontier under a fixed proposal model.

arXiv:2606.11221v1 Announce Type: new Abstract: We take a Gromov-Wasserstein perspective on Vision-Language-Action (VLA) learning, where the goal is to make the relational geometry of action representations compatible with the semantic geometry of VL embeddings. However, this alignment is non-trivial due to the mathematical heterogeneity between the domains: the semantic space of vision-language is topologically linear and isotropic, whereas the physical manifold of robotic action is non-Euclidean and anisotropic. Their disjoint metric structures render direct regression ill-posed. To resolve this incompatibility, we introduce LAST (Lie-algebraic Action Space Tokenizer), which reconstructs the action space to establish local metric compatibility with the VL modality via a two-stage transformation: (1) Global Topological Linearization: linearizing the action manifold via Lie-algebraic mapping, converting trajectories into a fixed-length, physically additive representation. (2) Local Metric Discretization: hierarchically discretizing the representation into schemas and whitened residuals, yielding approximately isotropic local charts that are statistically aligned with the semantic metric. By resolving the structural mismatch at both global and local levels, LAST enables VLA models with superior convergence and generalizability.

arXiv:2606.11233v1 Announce Type: new Abstract: Supervised Contrastive Learning (SupCon) has achieved strong performance by explicitly modeling pairwise relationships among samples. However, existing SupCon-based methods suffer from two key limitations: negative-sample dilution induced by the standard InfoNCE loss, and feature-space entanglement caused by the lack of explicit constraints separating category-relevant (common) and category-irrelevant (style) features. These limitations reduce feature discriminability and generalization ability. To address these issues, we propose OSCS-SupCon (Orthogonal Sigmoid-based Common and Style Supervised Contrastive Learning), a unified framework that combines a sigmoid-based pairwise contrastive objective with explicit orthogonality constraints. Specifically, we introduce a sigmoid-based contrastive loss with two learnable parameters, temperature and bias, which adaptively modulate pairwise decision boundaries and alleviate negative-sample dilution. Furthermore, we enforce orthogonality between common and style feature subspaces via a linear projection with ReLU nonlinearity, thereby reducing feature overlap and improving disentanglement of style-irrelevant representations. Extensive experiments on six benchmark datasets demonstrate that OSCS-SupCon consistently outperforms state-of-the-art supervised contrastive learning methods across multiple backbone architectures. In particular, on the fine-grained CUB200-2011 dataset with a ResNet-18 backbone, the proposed method achieves a 3.4% improvement in classification accuracy over CS-SupCon, highlighting its robustness and generalization capability. Ablation studies further confirm the effectiveness of each component.

arXiv:2606.11231v1 Announce Type: new Abstract: Vision-language reinforcement learning has recently shown strong target-present localization for camouflaged object detection (COD). Yet localization is only one side of the decision: when the agent faces an ordinary image with no camouflaged target, will it still claim that a camouflaged object exists? Standard COD training and evaluation data are positive-only, so agents optimized under this setting can acquire an over-detect bias, a task-specific form of object hallucination that standard COD evaluation leaves unmeasured. To quantify this target-absent behavior, we construct Counterfactual COD (CF-COD), a paired benchmark that removes the camouflaged target from each held-out COD evaluation image while preserving a plausible background. CF-COD evaluates whether a model detects the target on the original image and abstains on the target-absent counterfactual, summarized by Pair Accuracy (PA). We further introduce CFCamo, a paired counterfactual framework for COD with abstention. For training, CFCamo optimizes a Qwen3-VL-4B-Instruct agent with Counterfactual Sequence Policy Optimization (CSPO), which samples paired original-counterfactual rollouts and uses a Counterfactual Paired Reward (CPR) to couple original-image detection with counterfactual abstention. On CAMO-test, CFCamo improves S_alpha by +3.7 pp over the prior RL-based COD baseline; across CF-COD, it reaches 80.0-90.8% PA. Ablations show that removing counterfactual coupling reduces PA to 1.4-5.2% despite strong target-present COD scores, showing that target-present evaluation alone does not characterize detect-or-abstain behavior. Overall, these results indicate that CFCamo improves COD agents by coupling target-present detection with target-absent abstention, rather than merely strengthening target-present localization. Code and data are available at https://github.com/suhang2000/CFCamo.

arXiv:2606.11202v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed in applications for global multilingual users, yet safety training remains concentrated in dominant languages and has not progressed in parallel with multilingual capability, creating exploitable gaps for jailbreak attacks. Current jailbreak defenses are largely developed and evaluated in dominant languages, and their effectiveness is limited by the scarcity of aligned multilingual supervision and representations dispersion caused by language variation. To address this issue, we propose MLJailDe, a multilingual jailbreak detection framework designed to improve both multilingual robustness and cross-lingual generalization. MLJailDe first introduces a multilingual back-translation data augmentation algorithm to construct a semantically consistent and functionally effective dataset spanning 11 languages, consisting of 2,232 benign and 1,239 jailbreak samples. On this basis, MLJailDe employs relative-distance constraints to reduce cross-lingual representation dispersion and encourage jailbreak prompts with similar intent to form consistent clusters across languages, while an imbalance-aware classification objective is further used to alleviate class imbalance and learn more reliable multilingual decision boundaries. Experimental results show that MLJailDe outperforms state-of-the-art baselines across multiple languages, achieving an F1 score of 98.5\%, and obtains an average F1 score of 97.1\% on unseen languages, demonstrating strong effectiveness and cross-lingual generalization.

arXiv:2606.11285v1 Announce Type: new Abstract: Unauthorized unmanned aerial vehicle (UAV) activity around airports, public venues, and other sensitive sites has made protected-airspace monitoring increasingly important. A practical sensing system must search a wide angular region, find small long-range targets, and return both bearing support and UAV-specific evidence before a restricted perimeter is breached. Existing UAV detection paths often rely on spatially organized evidence, such as body extent, silhouette, or track continuity. At long range, however, these cues become difficult to preserve and verify as the target footprint weakens and its image-plane support shrinks. EventRadar follows a complementary cue: propeller-induced temporal periodicity, which recent event-camera sensing studies have shown can reveal UAV-specific motion after appearance becomes weak. We extend this cue to kilometer-scale active sensing with an event-camera prototype. Scene-Anchored Geometry Evidence (SAGE) fuses scanning events with IMU pose to maintain a bearing-indexed scene memory, separating transient candidate support from persistent background clutter. Comb-guided Harmonic-Group Learned Iterative Shrinkage and Thresholding Algorithm (CHG) then treats each candidate as a weak high-rate timing signal and recovers phase-insensitive harmonic evidence with fixed compute. Compared with related event-camera baselines on 700-1500 m UAV event recordings, EventRadar achieves 0.990 mAP$_{.3}$ and 0.949 F1$_{.3}$, reduces FN$_{.3}$ to 0.009, and shows real-time feasibility in prototype profiling.

arXiv:2606.11320v1 Announce Type: new Abstract: In this paper, we present a novel set of related models for semantic segmentation of node-link diagrams. These diagrams are frequently used to represent mathematical graphs, relationships between concepts, and flowcharts. Such diagrams are difficult to access non-visually; while some assistive interfaces have been designed for node-link diagrams, they rely upon a machine-readable representation of the diagram, whereas such diagrams will generally be made available as bitmap images. Our compact deep learning models show excellent quantitative and qualitative performance on a large synthetic dataset of node-link diagrams, reaching per-pixel accuracy over 93\%.