AllyshipContent Radar

Deep dives 738

Paper và bài viết kỹ thuật

Crawl live-1795068900000 · 13:22 19 thg 11, 2026 · Xuất bản 13:30 19 thg 11, 2026

Huấn luyện

DrawScore: Evaluating Generated CAD Models against Engineering Drawings

Automatic CAD generation is commonly evaluated with volumetric IoU, Chamfer Distance, and surface F-score, which barely penalize local errors that occupy only a small fraction of a part's volume or surface area, such as a misplaced hole or a slot on the wrong face, even though such errors are critical to manufacturing. We formalize these metrics in a unified integral form and show that the weight of a local error decays as the (k+q)-th power of its size, where k is the dimension of the integration domain and q is determined by the penalty function. Guided by this analysis, we propose DrawScore, which projects both models into three orthographic views with visible and hidden lines, as in engineering drawings, and applies a saturating penalty along the projected curves, so that the weight of a local error decays only linearly. Controlled experiments confirm the predicted decay orders, and on outputs of existing CAD generation agents, DrawScore exposes errors in outputs that IoU rates as nearly perfect. DrawScore further improves generation as a reinforcement learning reward and, for CAD generation from engineering drawings, serves as a test-time verifier without 3D ground truth or an additional reward model. Our code is available at https://github.com/UESTC-Simon/DrawScore.

  • DrawScore further improves generation as a reinforcement learning reward and, for CAD generation from engineering drawings, serves as a test-time verifier without 3D ground truth or an additional reward model.
Hugging Face Papers · Haorui Chen, Shiteng Lu, Zhanyu Guo, Zaixi Li, Yunhe Ni, Yaosong Du, Yichen Xiang, Xiande Huang, Rui Huang 2 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Let Samples Sing in Chorus: Mitigating Spurious Correlations via Cross-view Consensus

Deep models often rely on spurious correlations that fail on underrepresented groups. Existing methods either correct human-specified bias attributes or infer unknown biases from the learned representation space. However, annotated attributes may differ from the shortcuts actually used by a model, while a single representation provides incomplete and potentially noisy evidence. Our key insight is that a bias obscured in one feature space may be clearer in another. In this work, we propose CHORD (Cross-view Harmonic Outlier Reversal for Debiasing), a multi-view framework that allows samples to "sing in chorus" across diverse representation spaces to achieve robust cross-view consensus. CHORD aggregates aggregates within-class partitions in each view, which is a proxy for specifying candidate biases, to debias the target representation and classifier without bias supervision. Experiments on seven benchmarks across vision, language, speech, and medical imaging show consistent improvements in worst-group accuracy. Code and checkpoints are available at https://github.com/davelee-uestc/chord_debiasing_public .

  • In this work, we propose CHORD (Cross-view Harmonic Outlier Reversal for Debiasing), a multi-view framework that allows samples to "sing in chorus" across diverse representation spaces to achieve robust cross-view consensus.
  • Experiments on seven benchmarks across vision, language, speech, and medical imaging show consistent improvements in worst-group accuracy.
Hugging Face Papers · Weiwei Li, Junzhuo Liu, Tong Chu, Hengfu Yu, Yinjie Lei, Wen Li 2 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development

Games are a rich testbed for visual agents, demanding perception, long-horizon planning, and precise control. However, collecting human gameplay is costly, and learning from gameplay videos requires inferring missing actions. Existing datasets also cover a limited range of games, often without explicit tasks or verifiable outcomes. To address these gaps, we introduce Game2Player, a framework that co-develops game environments and players to collect diverse and verifiable gameplay experience at scale. In the Game2Player loop, a Builder agent adapts existing web games, exposes their internal state, and defines tasks with state-based progress measures and success checks. A Coding Player and a VLM Player then attempt the tasks and report problems to the Builder for repair, so that play serves as both environment validation and training data collection. With Game2Player, we build 568 verified web games with 3,436 tasks, collect 15.7K trajectories (307 hours of gameplay), and hold out 50 games as the benchmark PlayerBench. To evaluate the collected experience as training data, we train VLM agents with progress-weighted SFT, assigning greater weight to steps that advance task progress. Game2Player-9B improves progress from 14.3% to 31.2% on PlayerBench and from 30.0% to 53.1% on the external GameWorld benchmark. More broadly, our work suggests a path for visual agents to scale their gameplay experience by building, verifying, and learning from existing game worlds.

Hugging Face Papers · Xirui Jiang, Yuxi Chen, Rui Yang, Feng Luo, Pengcheng Wang, Bingxuan Li, Zihao Zhao, Yushan Li, Han Wang, Manling Li, Yang Zhang, Shiyu Chang, Tong Zhang, Huan Zhang 18 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

LogT: Logically Think with Images for Visual Search

Thinking with images empowers vision-language models (VLMs) to use an external Zoom-in tool for fine-grained visual search. Recent works employ supervised fine-tuning (SFT) and reinforcement learning (RL) with tool-specific designs to enable such capabilities. However, achieving logical reasoning alongside accurate and efficient tool use remains a significant challenge. To address this, we shift the focus away from increasing RL complexity and instead identify SFT data quality as a key bottleneck. We then propose LogT (short for Logically Think with Images), a fully automated SFT data engine that explicitly enforces logic-chain-guided reasoning. LogT generates an SFT dataset featuring challenging, visually grounded questions and structured tool-use traces with human-like logic. We validate our approach by fine-tuning two base models, Qwen2.5-VL-7B and InternVL3.5-8B, on the dataset and evaluating them across seven benchmark test sets. We show that LogT-SFT based on Qwen2.5-VL-7B outperforms the state-of-the-art trained with both SFT and RL, despite using 3.1times less data and >120times fewer GPU hours. Furthermore, applying a simple RL recipe without tool-specific design on LogT-SFT yields a model that surpasses prior methods with complex RL setups by a substantial 9.59-point margin. In addition to strong performance gains, our model exhibits higher tool-use accuracy and efficiency, as well as more structured and logically consistent reasoning traces, directly validating the effectiveness of our approach.

  • However, achieving logical reasoning alongside accurate and efficient tool use remains a significant challenge.
Hugging Face Papers · Yanjun Fu, Quanzeng You, Jiadong Guo, Yujie Lu, Sai Charitha Akula, Haiping Wu, Miao Liu, Sanghamitra Dutta, Lu Yuan 2 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

PIPETTE: Prompt Injection Purple-Teaming for Robust LLM Agents

Large language model (LLM) agents routinely interact with untrusted external data, exposing them to indirect prompt injection (IPI) attacks that hijack agents into executing attacker-chosen tasks. To uncover these vulnerabilities at scale, automated red-teaming methods often leverage reinforcement learning (RL) to generate attacks that achieve high attack success rates (ASR). Model developers use the attacks discovered during internal red-teaming to harden agents through adversarial training before deployment, a collaborative offensive-defensive process known in cybersecurity as purple-teaming. However, we show that achieving high ASR during red-teaming does not guarantee robustness gains from subsequent adversarial training. Specifically, we find that two state-of-the-art RL-based red-teaming methods, RL-Hammer and PISmith, suffer from mode collapse, repeatedly discovering similar attack strategies whose additional successes add no additional robustness improvements to adversarial training. To address this gap, we introduce PIPETTE, a unified prompt injection purple-teaming framework that effectively converts red-teaming discoveries into defender robustness improvements. PIPETTE employs a dictionary learning-based reward that encourages broader coverage of attack strategies during red-teaming. It also integrates an on-policy adversarial training pipeline tailored to IPI defense, enabling agents to resist attacks while completing legitimate user tasks. Extensive evaluations across agentic benchmarks (AgentDojo and ASB) demonstrate that PIPETTE produces defenders that achieve near-zero attack success rates on static attacks and can withstand strong adaptive attacks, while maintaining benign task utility close to that of the undefended baseline.

  • To uncover these vulnerabilities at scale, automated red-teaming methods often leverage reinforcement learning (RL) to generate attacks that achieve high attack success rates (ASR).
  • However, we show that achieving high ASR during red-teaming does not guarantee robustness gains from subsequent adversarial training.
  • Specifically, we find that two state-of-the-art RL-based red-teaming methods, RL-Hammer and PISmith, suffer from mode collapse, repeatedly discovering similar attack strategies whose additional successes add no additional robustness improvements to adversarial training.
Hugging Face Papers · Lipeng He, Yihan Wang, Jiawen Zhang, N. Asokan 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

A Foundation Model to Measure and Predict Human Behaviour

Large language models (LLMs) are increasingly used to infer psychological attributes, predict individual behaviour, and simulate human populations. However, despite generating plausible responses, they often misrepresent individuals and distort population statistics. Here we introduce Chiron, a foundation model to measure and predict human behaviour. We jointly fine-tune an LLM to represent psychological constructs, predict individual behaviour, and align population response distributions with human statistics. The training and evaluation data span 36 behavioural domains and 78 social science experiments, comprising 288,307 person records, 1,411,750 atomic behavioural decisions, and 99,953 psychometric assessments over 109 psychological constructs. In our evaluations, Chiron's psychological measurements agree more closely with human self-reports and exhibit fewer biases that undermine psychometric validity. It predicts individual behaviour more accurately than frontier LLMs (up to 28.4% relative improvement). In simulations of social science experiments, it more accurately captures the direction and magnitude of human experimental effects (82.2--94.3% lower effect-size estimation error). Together, these results establish Chiron as a unified computational instrument for measuring and predicting human behaviour.

  • 4% relative improvement).
  • Together, these results establish Chiron as a unified computational instrument for measuring and predicting human behaviour.
Hugging Face Papers · Haoran Ye, Tianrui Huo, Hangzhen Li, Xin Zhang, Guojie Song 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

UltraTex: Aligning Physical Knowledge for High-Fidelity PBR Texture Generation

Generating physically based rendering (PBR) texture maps for a 3D mesh from a single reference image is highly ambiguous, as the observed appearance entangles material identity, surface reflectance, and illumination. While current generative methods produce realistic albedo maps, they still struggle to infer accurate metallic and roughness attributes for relighting. We present UltraTex, a cascaded multi-view framework that moves beyond appearance matching by integrating physical knowledge into multi-view diffusion-based PBR generation. UltraTex first jointly predicts albedo and world-space shading normals to generate more detailed albedo maps, and then predicts metallic-roughness (MR) maps conditioned on the generated albedo, reference image, and material semantic priors. To inject material semantics, a lightweight adapter transforms hidden-layer features of a vision-language model (VLM) into intrinsic material representation tokens, providing part-level knowledge anchors for MR estimation. To model material-light interaction, photometric representation alignment guides intermediate diffusion features toward ground-truth material renderings under controlled lighting. This makes metallic and roughness variations observable through highlight intensity, specular sharpness, and local reflectance contrast, providing stronger guidance for fine-grained material prediction without explicit lighting supervision. Extensive experiments show that UltraTex generates high-quality textures and outperforms existing open-source and commercial methods in visual fidelity, material accuracy, and physical plausibility.

  • Extensive experiments show that UltraTex generates high-quality textures and outperforms existing open-source and commercial methods in visual fidelity, material accuracy, and physical plausibility.
Hugging Face Papers · Jingzhi Bao, Xiaohe Ma, Hong Li, Yuanming Hu, Xiaoguang Han 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Show HN: KradleBench – Multi-agent AI evals in Minecraft

Hi HN - This is James and Kemal. In past lives we led Firebase and Tensorflow/JAX respectively. HN has been a particularly big part of my (James’) career and the trajectory I’ve taken would be meaningfully different without it. Thank you! Our team shipped a benchmark: KradleBench. We test AI in Minecraft Worlds. 1. The worlds are multi-agent. Humans and other AIs are in there too, often pursuing their own goals. Collaboration and conflict skills fall out of this property. 2. The world are dynamic. They constantly change, which means that agents need to react. “Capture the Flag” has a single winning condition, and many different ways to achieve it. This leads to many emergent and creative strategies. 3. The worlds are real-time. They are not turn-based, they do not wait for AIs to finish thinking. Nature balanced speed and intelligence in humans, we can understand this in AIs too. Underneath all this is Kradle, a platform to test and understand intelligence. It’s a petri dish where any intelligence type (LLM, World Model, Human) can be compared and generate useful information to improve current systems, whether it’s on capabilities or safety. Some findings so far: - A skilled human was #1, until yesterday when Astra 6.2 surpassed them. We lose badly on solo tasks and out-compete on multi-agent ones - Astra 6.2 iterates on its strategy 4x faster than Fable 5.2. The Claude family tends to think hard and stick with strategies. The OpenAI models tend to be more reactive and iterative, which serves them well in multi-agent, dynamic, real-time evals. We’ve got videos on all the top model’s page, and we’ll be producing them when a new model launches. Sign up on the site to get them and for when the platform opens. Thanks for reading, James & Kemal

  • “Capture the Flag” has a single winning condition, and many different ways to achieve it.
  • It’s a petri dish where any intelligence type (LLM, World Model, Human) can be compared and generate useful information to improve current systems, whether it’s on capabilities or safety.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Out of Sight, Still in Mind: Token Compression for Omni-LLMs

The goal of this paper is to reduce the input token cost of Omni-modal large language models (Omni-LLMs) at inference time. Omni-LLMs reason jointly over audio, video and text, but the cost of the three streams is highly unbalanced: visual tokens account for the vast majority of the input, and are highly redundant. In this paper, we propose ReMo, a training-free framework that compresses visual tokens by redistributing their information across modalities: a visual token is kept only if its information appears nowhere else. ReMo achieves this in two ways: (i) it aligns audio and video in a shared space via closed-form calibration, and removes visual tokens already explained by the audio or by other visual tokens; and (ii) it replaces object-level visual tokens, identified by a lightweight object detector, with compact text proxies, short descriptions of each object and its location, conveying the same content in far fewer tokens. On Qwen2.5-Omni at two model scales, ReMo removes 54% of the input tokens with comparable accuracy. It performs comparably to the full-token model, reaching 101.2% and 101.4% of its average accuracy over five audio-visual benchmarks.

  • The goal of this paper is to reduce the input token cost of Omni-modal large language models (Omni-LLMs) at inference time.
  • ReMo achieves this in two ways: (i) it aligns audio and video in a shared space via closed-form calibration, and removes visual tokens already explained by the audio or by other visual tokens; and (ii) it replaces object-level visual tokens, identified by a lightweight object detector, with compact text proxies, short descriptions of each object and its location, conveying the same content in far fewer tokens.
Hugging Face Papers · Suho Yoo, Youngjoon Jang, Hyebin Cho, Joon Son Chung 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

NexSight: Verifiable Tasks for Fine-Grained Agentic Long-Video Understanding

Fine-grained questions about long videos ask how many times an action occurs, in what order events unfold, or exactly when a state changes. Answering them can require finding brief moments scattered across hours of footage. Agentic models that decide when and where to look promise to find these moments, yet this ability is hard both to measure and to learn. Sparse supporting evidence may establish that an action occurred without establishing its total count; verifying timeline-wide claims requires checking for missed occurrences and double counting. Answer-only supervision, in turn, does not explicitly supervise the search for evidence. We present NexSight, which builds verifiable tasks to both evaluate and train agentic long-video understanding. Its agentic pipeline, SightForge, accepts a task only if a verifier blind to the draft answer can reconstruct it from evidence densely reacquired from the raw video and counterexample checks fail to refute it. This yields SightLine, 310 evidence-traceable tasks from 108 long videos, evaluated through SightLoop, a shared multiscale harness supporting temporal exploration, local inspection, evidence integration, and revisiting while recording returned observations. On disjoint videos, solver trajectories on verified tasks become step-level supervision for observation decisions. The strongest model reaches only 54.52% on SightLine, yet 79.35% when given the annotated evidence. With under 6K examples, Sight-35B nearly doubles its base model's interactive accuracy (21.29% to 42.26%), outperforms the 397B model of its family, begins using explicit spatial inspection after post-training, which its base model never uses, and gains over 8 points on two external long-video benchmarks.

  • 26%), outperforms the 397B model of its family, begins using explicit spatial inspection after post-training, which its base model never uses, and gains over 8 points on two external long-video benchmarks.
Hugging Face Papers · Tianci Xie, Li Chen, Pinxue Guo, Rui Zheng 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

VAE Rhythm: Temporal Decoding Traces for AI-Generated Video Detection

Visible artifacts are an unreliable cue for detecting generated videos, because newer generators learn to remove them. We instead detect a trace left by the generation pipeline itself. Most video generators decode latents with a 3D causal VAE that groups consecutive frames in time, so the fidelity of generated frames rises and falls periodically with their position in a group. We call this periodic trace VAE Rhythm. Real videos are captured frame by frame and do not have it. To reveal it without training, we propose a phase scan. The scan reconstructs a clip with an off-the-shelf VAE from successive start offsets and measures the periodicity of the reconstruction error across these offsets. We find VAE Rhythm in most of the 39 variants of 31 generators we test, including the closed-source Sora, Kling 3.0 and Seedance 2.0. We use the rhythm as a zero-shot score, as a plug-in for existing detectors and as a data engine that pairs real videos with their VAE reconstructions to train our detector CaDet. On DF26, the training-free rhythm score outperforms all 15 retrained baselines, and as a plug-in it improves all 19 tested detectors by 6.7% to 23.7% AUROC. CaDet achieves the best AUROC on AIGVDBench, DF26 and three 2026 closed-source models, and its variant trained without any generated video still exceeds 90% on AIGVDBench.

  • CaDet achieves the best AUROC on AIGVDBench, DF26 and three 2026 closed-source models, and its variant trained without any generated video still exceeds 90% on AIGVDBench.
Hugging Face Papers · Ke Sun, Shuwei Li, Haiyang Mei, Mike Zheng Shou 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Tell HN: Notion AI chat sessions have been exposed via Fireworks AI hack

Only an email so far, nothing official I could find. I figure if you didn't get this email you are ok. --- We are writing to inform you of a security incident at one of our vendors, Fireworks AI, that may impact you. For the latest on the Fireworks AI incident, see their official updates here. Following an investigation by the Notion security team, we have determined that your workspace, [...], was among the <1% of Notion workspaces that may have been affected by the Fireworks security incident. Some unattributable chat sessions with Notion AI may have been exposed in this incident. “Unattributable” means the chats weren't labeled with who wrote them or which workspace they came from, and the transcripts did not contain tool call results. Beyond what’s in the chat transcript itself, there’s no information tying the chat back to your workspace, users, or company. No other data from within your workspace is affected. This email explains what happened, what data of yours might be affected, what we’re doing to address this, and what we recommend to protect your data from this Fireworks AI incident. What is Fireworks AI and how we used them: Fireworks is one of the most used global AI infrastructure platforms powering tens of thousands of customers and used by hundreds of thousands of developers. We used Fireworks for classification jobs to help customers understand their own AI usage through a feature in Notion called Insights, and we used them as a general inference provider to serve open-weight models inside our product. The Insights feature has been disabled as we investigate this incident, and we are no longer serving open-weight models through Fireworks. What happened and how we addressed it: We learned from Fireworks about potential incidents earlier this month, and we shut down all connections to their systems out of an abundance of caution. On November 6, Fireworks notified us that malicious actors may have accessed potential customer data. Our team immediately began another internal investigation. On November 8, Fireworks confirmed this new dataset for us. Upon analysis of the new data, we discovered that it included unattributable Notion AI chat transcripts used for the Insights feature. Chats weren't labeled with who wrote them or which workspace they came from, and the transcripts did not contain any tool call results. Beyond the words in the chat transcripts themselves, the data contained no information that could specifically identify your workspace, users, or company. No other Notion data from your workspace is affected, and we continue to monitor this incident. Recommended next steps for you: You can retrieve and review the impacted data securely inside your workspace. See below: - [...] We recommend that you treat this data as if it were exposed. If anyone on your team might have pasted sensitive information into chats in Notion AI like passwords, API keys, or other credentials during this timeframe, please rotate them now. You trust us with your life's work in Notion, and that trust covers every partner we choose. In this critical time for global security, it is our mission to raise the standard for how your data is protected. You have our uncompromising commitment. --- I downloaded the dataset and it's a full chat for me.

  • “Unattributable” means the chats weren't labeled with who wrote them or which workspace they came from, and the transcripts did not contain tool call results.
  • Chats weren't labeled with who wrote them or which workspace they came from, and the transcripts did not contain any tool call results.
HN: open weights · RamblingCTO 19 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Show HN: One cloud computer for all your agents

Hi HN, my team just released box by Simular, a cloud computer for all your agents, with fast computer use built-in. box gives your agents one remote Ubuntu desktop with Chrome and a shell, which your AI can then drive over MCP while your own computer stays unoccupied. That means you can turn your basic ChatGPT plan into a personal agent by giving it a box computer; let your Claude Code pick up what Codex has left off on the same box to maximize usage; actually close your laptop and let tasks run in one persistent, 24/7 environment with all your logins and setup. Some box demos: https://x.com/SimularAI/status/2121329778896871665?a=U4U1yYT... We started on this because we believe computers shouldn't need human operation in the future. When agents arrived, people started running them on mini computers and you see colleagues walk around the office with their laptop lid open. We think there's a better solution - cloud computers engineered for agents. Simular started three years ago to build out the full stack of an autonomous computer. box is our attempt to give agents a body to live in. Our computer use outperformed Opus and GPT on OSWorld, costing fewer tokens. https://thenewstack.io/sai-agent-osworld-benchmark/ Copy/paste one command into your agent to start: "Set up box from https://box.simular.ai/install " You can watch box's live action from a dashboard https://box.simular.ai/box We'd love for you to try it and share your feedback. I'll be in comments here to answer any question.

  • Our computer use outperformed Opus and GPT on OSWorld, costing fewer tokens.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

WorldRuler: Interactive World Modeling under Configurable Rules

Rules govern how player actions change interactive worlds. Different rules create distinct gameplay experiences within similar scenes and action spaces. For customizable gameplay, generative world models must follow user-defined rules beyond real-world behaviors. However, realizing user-defined rules presents two challenges: Rule-State Persistence and Rule-to-Visual Realization. The former requires preserving rule definitions, dependencies, and world states across saving and loading, so that restored worlds follow the established rules. The latter requires accurately realizing rule-induced changes on designated objects and attributes, particularly when affected objects are unseen. In this paper, we propose WorldRuler, which organizes rule definitions, dependencies, and object states in an executable 3D rule graph and connects rule execution, video generation, and perceptual feedback in a closed loop. The graph supports joint saving and restoration of rules and world states, while dynamic object binding grounds rule-specified entities in visual instances and uses stored states to constrain generation from their first appearance. We further introduce object attribute state replacement and counterfactual reconstruction training, enabling accurate attribute control and compatible appearance and motion changes while limiting interference with unrelated content. We develop an automated data collection and annotation pipeline using GTA-V and Habitat to construct RuleVerse, a dataset of rule-driven interactions with detailed state and event annotations, including cases where rules take effect before affected objects become visible. Extensive experiments demonstrate the effectiveness of WorldRuler in preserving rule-state consistency across saving and loading and accurately realizing rule outcomes on designated objects and attributes.

Hugging Face Papers · Mingliang Zhai, Chuanhao Li, Jiaming Tan, Zhen Li, Yuwei Wu, Yunde Jia, Zian Meng, Yuanyang Yin, Kaipeng Zhang 8 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Code World Model 2.0: Towards Infinite Controllable Video Generation

Visual proxies decouple explicit world-state simulation from photorealistic video synthesis, offering a controllable interface for both offline clip creation and real-time interactive streaming. Yet a practical renderer must address two challenges. Upstream proxies vary from coarse layouts to fine articulated geometry depending on the interaction, and streaming inference must operate in a few causal steps without accumulating drift across self-generated history. We present Invoker-Render, pairing a bidirectional creation model with a causal four-step streaming model under a shared conditioning interface over structural proxies, an appearance reference, and language. Trained on an aligned approximately 200-hour multi-granularity corpus, Invoker-Render honors fine geometric constraints when supplied while synthesizing plausible local motion under coarse proxies, enabling runtime selection of input detail. To bridge bidirectional creation and real-time streaming without compounding step compression and rollout drift, we introduce Progressive Forcing (ProForcing). ProForcing first trains a causal predictor and distills it into a four-step renderer using histories derived from training videos. Training then shifts to continuous self-generated rollouts. A pipelined serving stack supports efficient streaming inference. Experiments show versatile control across proxy families, granularities, and camera paths, consistent gains over recent streaming baselines, and stable 30-minute racing rollouts that preserve scene and subject appearance.

Hugging Face Papers · Ruoyu Wang, Hao Zhong, Yanming Yang, Jiahao Tian, Lin Liu, Biao Gong, Guoshan Lu, Yiding Tian, Chunhua Shen, Zhenzhong Lan, Chi Zhang 8 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

How are large companies not ripshit about the frontier labs?

Câu hỏi Can anyone name one positive thing that's come from AI so far objectively?

I am utterly baffled that the entire world is not coming after Anthropic and OpenAI with pitchforks. We are essentially funding: 1.The largest intellectual property theft from humanity in history, including corporate IP 2. Outrageous amount of money being spent with no real upside to show at this point. Can anyone name one positive thing that's come from AI so far objectively? 3. Non stop threats and attacks and greater and greater risk of bad actors engaging in cyber attacks. 4. Untold disruption to the economy and jobs 5. Untold risk from bioweapons 6. Increased instability in multiple sectors 7. Unknown risks to human intellect, development, and emotional well being I could go on. I just don't understand how large companies and governments all over the world aren't beside themselves with anger and rage. I could understand if we were seeing tangible improvements to work, society, and the products we use on a daily basis, but as someone who works for a very very large company who is embracing AI, we have seen almost 0 net benefit, with most people feeling burnt out, unhappy, scared, and overworked.

  • I could understand if we were seeing tangible improvements to work, society, and the products we use on a daily basis, but as someone who works for a very very large company who is embracing AI, we have seen almost 0 net benefit, with most people feeling burnt out, unhappy, scared, and overworked.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Show HN: A deterministic typesetting tool for Orthodox Music

I'm a trained Russian Orthodox church musician and software developer. I figured I would flog my Claude Pro subscription into helping me create a deterministic engine for typesetting church music that would otherwise be sung off of text alone in most cases. No AI is used in processing any of the texts, it's entirely algorithmic. Even if you don't have a use case for this, I figured the music nerds out there might appreciate taking a look at it and even perhaps giving a suggestion on how to improve it.

  • Even if you don't have a use case for this, I figured the music nerds out there might appreciate taking a look at it and even perhaps giving a suggestion on how to improve it.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Show HN: Hob - Open-source, on device alternative to Claude Cowork and ChatGPT

Hey HN, I'm Christian and this is a new OSS project I've been noodling on for the last couple of weeks. Hob is a free, MIT-licensed desktop app for Mac, Windows and Linux that works on your files the way the Claude and ChatGPT desktop apps do, except the model runs on your own computer. It uses open-weight models (currently Gemma 4 and gpt-oss). All your chats and files never leave your system. The only things that go online are web search (can be turned off in settings) and any MCP connectors you add yourself. Main features: - Automatically finds the best open-weight model for your hardware on first install (Mac/Linux/Windows) - Writes sandboxed Python code for ad hoc tools (data processing, document conversion, graphs etc.) - Built-in web search (powered by SearXNG) - MCP support - Image recognition - CRUD to your local filesystem - Works fully offline with no internet Before building this I tried every local AI app I could find, but it seemed each one had different problems. The main thing I noticed is that none of them could do what the ChatGPT and Claude apps do with your files, e.g. read and change files in a folder, look at images, run code in a sandbox, draw a chart. The most advanced thing any of them did was paste a plain text file into a small context window, which basically failed as soon as you had a file larger than a few KB. I couldn't find anything that could easily parse images and bigger PDFs. The OSS solutions that were more advanced all seemed to default to hosted models (either OSS or closed), which defeats the entire point of the privacy that on-device OSS offers. They also seemed to push towards an account/open-core style business. I build it using Electron bridged with llama.cpp in its own process. On initial load, the system will try and figure out if you have a compatible graphics card and if not default to CPU-only (I ran into a lot of issues of graphics card beeing seemingly compatible but then crashing on first query, there's a whole saga to this lol). Code execution happens in a sandbox that can only write to the chat's folder and has no network unless asked (sandbox-exec on macOS, Landlock on Linux and a low-integrity process on Windows). Images go to a small Qwen3-VL model. Dictation is a local Whisper model. The hardest part was definitely to get the smaller models to use tools reliably. Ran into a ton of issues of 4B models happily saying "here's your chart" without ever calling the chart tool or get stuck in an infinite call loop. I spent a fair amount of the work building guards vs tweaking prompts (that didn't really work for small models). I understand the pushback that local models are weaker than frontier cloud models. This is true, but for most tasks for most normal people, I'd argue you rarely need the newest frontier model. From my own evals, systems with 16+ GB of RAM do really well, and if you have a GPU or more RAM, you get very close to what you'll get out of the box with a frontier hosted model. Excited to hear what you think and see how it runs on other systems!

HN: open weights · pancomplex 180 lượt bình chọn 44 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Jeveration: System One Models Now Generate

System One models such as Jev return decisions through fast, parallel prediction. In this paper, we introduce Jeveration, showing that the same decision paradigm can also generate high-quality images in a single forward pass. Jeveration builds on a frozen vision-language model, repurposing its visual-semantic knowledge for holistic image generation. The key challenge is to maintain global consistency. A single prompt admits many valid images, yet parallel predictions must converge on a coherent whole. We therefore condition all positions on a shared random state to coordinate parallel predictions, so that each draw yields a different but coherent image. Since the model tends to ignore this state under direct image supervision, we first learn a distribution over discrete image codes to guide single-pass generation. Beyond text-to-image generation, we further unify image-to-image generation and dense prediction within the same architecture, so that one checkpoint supports all three tasks, each in one forward pass. Experiments show that Jeveration generates a 512x512 image in 78 ms, up to 122x faster than existing unified models, while maintaining competitive image quality and dense prediction performance.

Hugging Face Papers · Yuhe Liu, Qi Li, Shihua Zhang, Xinchao Wang 5 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Training Language Models for Introspective Oversight

arXiv:2611.23365v1 Announce Type: new Abstract: Language models increasingly exhibit behaviors that are difficult to detect in transcripts, such as reward hacking, deception, and sycophancy. Asking the assistant follow-up questions about its actions is unreliable, since models are generally not trained to report their own misbehavior. As a result, approaches to oversight typically involve querying external models. These face limitations: external transcript monitoring operates only on visible model outputs, and activation decoders lack full context and are prone to confabulation. We introduce introspective oversight, which trains a new role in the model to answer arbitrary questions about the assistant, rewarding answers that are verified against the assistant's actual behavior. We implement this by adding a separate, third-person observer role that answers these queries, which we can train solely for accuracy and separately from the assistant. On five held-out oversight evaluations, an observer trained on Qwen3.6-27B obtains a mean AUROC of 0.907, outperforming honesty training (0.864), transcript monitoring (0.828), and activation oracles (0.796). We further find that training an observer makes the assistant itself easier to monitor, with every baseline method improving on average when applied to the fine-tuned assistant. These results suggest that training models to answer verifiable questions about themselves is a tractable approach to overseeing unintended behavior.

  • As a result, approaches to oversight typically involve querying external models.
  • We further find that training an observer makes the assistant itself easier to monitor, with every baseline method improving on average when applied to the fine-tuned assistant.
  • These results suggest that training models to answer verifiable questions about themselves is a tractable approach to overseeing unintended behavior.
arXiv cs.AI · Alex Lin Wang, Linus Yifeng Tang, Dami Choi, Neil Chowdhury Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
21–40 / 738
Trang 2 / 37