Tuesday, July 21, 2026

1. How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs?
   T2arXiv AI/CL/LG（论文）

   Modern large language models (LLMs) are alarmingly susceptible to surprisingly simple immaterial changes in input prompts, such as casual hints, mislabeled few-shot examples, or fake prior assistant turns, which often flip originally correct answers. This paper systematically investigates where this susceptibility, spanning sycophancy and related cue-induced biases, resides inside the model. Across five model families and seven BCT bias types, the authors extract per-bias directions from hidden states and triangulate them through probing, leave-one-dataset-out transfer, and causal intervention. The key finding is that this susceptibility is largely installed by alignment tuning rather than pretraining. Within aligned models, each bias becomes a single coherent direction that can be decoded and steered along to recover unbiased answers. These biases remain representationally distinct, with cross-bias entanglement being model-specific rather than a property of the bias category. The same intervention also serves as a modest debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers.

2. The Calibration Channel Determines the Bayes-Error Proxy: An Exact Law for Temperature-Induced Distortion
   T2arXiv AI/CL/LG（论文）

   This paper precisely characterizes how temperature scaling—the most widely used post-hoc calibration method—distorts the soft-label Bayes-error proxy. The authors prove a model-free identity reducing the temperature-scaled proxy to the classifier's margin distribution, yielding strict monotonicity in temperature and a continuous bijection from temperature to (0,1/2), meaning a fixed classifier can report any proxy value. On CIFAR-10, Fashion-MNIST, and SVHN, the proxy varies 56x to 980x at constant test error, and the calibration temperature minimizing expected calibration error does not coincide with any stable proxy value. The results reinforce that a proxy value is meaningful only together with the mechanism that produced its probabilities.

3. Google is Developing a New AI Chip to Make Gemini Models More Efficient
   T2TechCrunch：AI（RSS）

   Alphabet's Google is reportedly developing a new AI chip specifically designed to significantly improve the computational efficiency of its Gemini models. This move aims to reduce running costs and accelerate inference speed, further solidifying Google's leadership in AI. The chip is expected to be deployed in data centers and edge devices, providing a more efficient hardware foundation for the Gemini series.

4. AI's Most Important Protocol Gets Easier to Use
   T2TechCrunch：AI（RSS）

   The Model Context Protocol (MCP), a foundational building block for AI interoperability, is getting usability improvements. The update lowers the integration barrier for developers, enabling AI models to securely access external data sources, potentially accelerating data exchange between AI applications and external systems.

5. Causal Discovery on Irregular Time Series
   T2arXiv AI/CL/LG（论文）

   This paper proposes an extension of PCMCI+ for causal discovery on irregularly sampled time series. Traditional methods rely on fixed-lag structures, limiting applicability to irregular streams such as sensor data, healthcare records, and financial transactions. The new approach aggregates causal influences over predefined temporal windows, consistently recovering the true causal graph and substantially outperforming standard PCMCI+ on synthetic irregular data.

6. Google DeepMind Introduces Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
   T1Google DeepMind：Blog（RSS）

   Google DeepMind announced three new Gemini models on its official blog: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. These models continue the Flash series' focus on efficiency and lightweight design, offering developers and enterprises more flexible options between performance and cost. The 3.5 Flash-Lite further lowers computational requirements, while the 3.5 Flash Cyber may be optimized for cybersecurity or specific domains. This release underscores Google's continued investment in lightweight large language models.

1. Grabette: An Open System for Recording Robot Manipulation Data
   T1Hugging Face：Blog（RSS）

   Hugging Face team releases Grabette, an open system designed to streamline the recording and collection of robot manipulation data. It supports multiple sensor inputs, provides standardized data formats, and integrates visualization and replay tools to help researchers efficiently build robot learning datasets. Grabette's modular design makes it easy to extend and compatible with mainstream robot platforms, potentially advancing open and reproducible embodied AI research.

2. Claude Code v2.1.216: Fixes Long Session Performance and Numerous Stability Issues
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code v2.1.216 fixes a quadratic message normalization cost that caused multi-second stalls in long sessions, improving response speed for extended use. It resolves critical issues such as OAuth token expiration causing auto-mode command denial, worktree-isolated subagents reverting to default agent, @-mentions failing after file-modifying hooks, and background session cancellations. Improvements include a compact /fork confirmation line, explicit /context window overflow warning, better /spend-limit adjustment prompt, and /code-review empty-diff message. Additional fixes cover Windows permission prompts, PowerShell argument validation, fullscreen dialog overflow, and right-to-left text rendering.

3. One Person Runs 200 AI Agents: Practical Experience in an AI Agent MMO
   T2Hacker News 热门（Algolia · 中文化）

   In SpaceMolt, an MMO exclusively for AI agents, Brocktree operates approximately 200 AI agents using a single stationary hub bot. He advocates distrusting agents' planning capabilities and instead using lightweight scripts as a cheaper alternative to expensive token calls. This counterintuitive approach highlights the trade-off between cost and reliability in large-scale AI agent operations, offering valuable insights for multi-agent system deployment.

4. Nativ: Run AI Models Locally on Your Mac
   T2Simon Willison 博客（Atom）

   Nativ is a new project by developer Prince Canuma, built on MLX to provide a full macOS desktop application for running AI models locally. It features a chat interface and a localhost API server, and automatically detects models already cached from Hugging Face, similar to LM Studio but optimized for Apple Silicon. This project extends the author's previous excellent MLX-VLM library, further lowering the barrier to using local LLMs.

5. Gritt exits stealth with $34 million for robots to build solar plants—then, everything else
   T2TechCrunch：AI（RSS）

   Gritt is coming out of stealth with $34 million in funding, aiming to automate the hardest tasks on construction sites using robots. The initial focus is on building solar plants—handling heavy lifting, assembly, and other labor-intensive work—with plans to expand into broader construction. This addresses labor shortages and improves efficiency and safety in the industry. Backed by top venture firms, the company could accelerate renewable energy infrastructure deployment.

6. Reverse Engineering Is Now Cheap: Coding Agents Lower the Barrier for Hardware Automation
   T2Simon Willison 博客（Atom）

   Simon Willison observes a growing number of people using coding agents to reverse-engineer and automate home devices. Previously, the effort was rarely worthwhile due to high maintenance costs and unstable APIs. Now, coding agents drastically reduce the cost of trying and failing, making home automation practical and easing the psychological burden of future maintenance. This shift redefines the ROI calculation for developers and could spark a wave of customized device hacking.

1. Anthropic Launches Rare Disease Research Grants for AI for Science, Offering Up to $50,000 in Claude Credits
   T1Anthropic：News（网页）

   Anthropic has announced a focused call for AI for Science applications targeting rare genetic diseases. Selected applicants will receive up to $50,000 in Claude credits over six months, aiming to build a community of researchers exploring how AI can reshape our understanding of rare diseases. The program is open globally and encourages innovative use of Claude for data analysis, gene function prediction, and more.

2. Deezer Says Over 50% of Daily Music Uploads Are AI-Generated
   T2TechCrunch：AI（RSS）

   Music streaming platform Deezer reported in July 2026 that over 50% of daily music uploads on its platform are AI-generated, with more than 90,000 AI tracks uploaded per day in June. This raises concerns about content quality, copyright protection, and moderation mechanisms, signaling how AI-generated content is rapidly transforming the digital music landscape.

3. Anthropic's $1.5B Copyright Settlement Gets Final Approval
   T2TechCrunch：AI（RSS）

   A US court has given final approval to Anthropic's $1.5 billion copyright settlement with a group of authors, ending that specific lawsuit. However, the case does not resolve the broader legal issue of using copyrighted works to train AI models, with the industry awaiting further precedents and legislation.

4. Who's Afraid of Chinese AI Models? US Should Legalize Distillation and Fair Use
   T2Simon Willison 博客（Atom）

   Ben Thompson proposes a bold policy: the US should pass a law making training data collection fair use and barring terms of service that forbid distillation, to help US open models compete with Chinese counterparts. He argues distillation is nearly impossible to stop, so the US should adopt a new copyright policy that indemnifies labs and lets their learnings fuel further innovation. Meanwhile, Alibaba's decision to release Qwen 3.8 Max as open weights may have been influenced by Xi Jinping's speech encouraging open source and collaboration.

5. OpenAI Fears Open-Weight Models: Should the US Ban Chinese-Made Ones?
   T2TechCrunch：AI（RSS）

   TechCrunch reports that OpenAI is privately lobbying the US government to consider restricting Chinese-made open-weight LLMs, arguing they could undermine US AI leadership. However, critics question: Would banning open-source models stifle innovation? Turning AI into a sustainable business has become the central tension, and the open-vs-closed model debate has evolved from a technical issue into a geopolitical one.

6. Trump's Latest AI Czar Has Already Resigned
   T2TechCrunch：AI（RSS）

   The director role for the Center for AI Standards and Innovation (CAISI) has become a revolving door since David Sacks left his position as AI czar. Trump's latest appointee has already resigned, highlighting instability in U.S. AI policy leadership and adding uncertainty to global AI governance.

1. Automated Discovery Has No Universally Superior Harness
   T2arXiv AI/CL/LG（论文）

   This paper systematically decomposes the components of automated discovery harnesses like OpenEvolve and TTT-Discover, and evaluates 30 budget-matched harnesses across 12 model-problem pairs using over 3.1 million LLM rollouts with repeated-trial statistical analysis. The results show no fixed harness is universally superior, and OpenEvolve variants generally underperform simpler alternatives. The authors suggest treating harness choice as a hyperparameter rather than a universal recipe, and propose an adaptive allocation strategy based on early discovery progress. All run pools and baseline null distributions are released as reusable statistical infrastructure.

2. Certified Training for Convolutional Perturbations: Formal Safety Guarantees for Vision Models
   T2arXiv AI/CL/LG（论文）

   Vision models are highly susceptible to convolutional perturbations like motion blur from camera shake, risking failures in critical applications. While data augmentation and adversarial training improve empirical robustness, they lack formal guarantees. This paper introduces a novel certified training approach that efficiently encodes convolutional perturbations to train provably robust models. On CIFAR10, it achieves over 80% robust accuracy against moderate motion blur while maintaining comparable standard accuracy, significantly outperforming adversarial training.

3. FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications
   T2arXiv AI/CL/LG（论文）

   FlashRT is an agent harness that guides generic coding agents to transform simple reference implementations into optimized multi-GPU deployments via multi-pass transformation, intermediate representation validation, and measurement-gated optimization loops. It achieves up to 70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs, and on AMD MI355X GPUs matches latency reduction while increasing throughput improvement to 3.6x, with 65% latency reduction for Qwen3-Omni text-to-audio inference compared to the expert vLLM-Omni implementation.

4. The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric
   T2arXiv AI/CL/LG（论文）

   Human visual similarity judgments are highly context-dependent—two images may be similar in shape but distinct in color, yet existing metrics collapse these nuances into a single scalar without conditioning on specific aspects. This work introduces a large-scale dataset of human similarity judgments over image triplets, each annotated across multiple free-form semantic aspects. Benchmarking frontier Vision-Language Models (VLMs) reveals a substantial gap compared to human consensus. By fine-tuning a VLM on this data, the authors propose the Text-Prompted Image Perceptual Similarity (TPIPS) metric, which captures multiple senses of visual similarity based on a text prompt. TPIPS aligns better with human perception and generalizes reliably beyond the training distribution. It unlocks new capabilities in text-guided retrieval, compositional search, and fine-grained evaluation of generative models.

5. Patch Policy: Efficient Embodied Control via Dense Visual Representations
   T2arXiv AI/CL/LG（论文）

   Patch Policy is a lightweight Transformer policy architecture that directly leverages pretrained dense patch features from Vision Transformers without the need for a full large vision-language model. By using a block-causal attention mask, it preserves temporal causality while allowing the model to attend over multiple patch tokens per observation along with other state information. Across 4 simulated and 3 real-world environment suites, it achieves a 40% relative improvement over global-pooled policies and surpasses fine-tuned OpenVLA-OFT by 18% using only 0.7% of the parameters. This provides the robotics community with a pipeline to efficiently leverage progress in visual representation learning.

6. Do Language Models Dream of Binding Molecules? Benchmarking LLMs under Spatial Constraints
   T2arXiv AI/CL/LG（论文）

   This paper systematically analyzes the ability of general-purpose large language models (LLMs) to handle 3D spatial constraints in structure-based drug design (SBDD). Compared to specialized models like diffusion models, LLMs still lag behind in multi-conditioned spatial molecule generation tasks such as protein pocket constraints, pharmacophore points, and mandatory interactions, but show promise in handling multiple spatial constraints simultaneously. The authors introduce 3D-Fit, a token-efficient benchmarking strategy, to evaluate LLMs and reveal their scalability to heterogeneous setups.

7. Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering
   T2arXiv AI/CL/LG（论文）

   This paper analyzes self-loop failure trajectories in LLM reasoning, and proposes SOPHIA, a method that treats reasoning traces as sequences of latent states. By classifying prefixes into states, recording transitions, and constructing a bank of steering vectors indexed by state pairs, a controller can detect self-loops online and apply targeted interventions. Experiments show that SOPHIA reliably prevents self-loops, improving end-task accuracy and token efficiency, demonstrating fine-grained controllability over reasoning processes.

8. GigaPath-Flash and GigaTIME-Flash: Efficient Pathology Foundation Models for Whole-Slide and Tumor Microenvironment Analysis
   T2arXiv AI/CL/LG（论文）

   This paper introduces GigaPath-Flash and GigaTIME-Flash, two efficient pathology foundation models. GigaPath-Flash distills a 22M-parameter tile encoder and a 21M-parameter slide encoder from the billion-parameter GigaPath teacher via knowledge distillation, retaining 97% of slide-level performance with 50× less compute. GigaTIME-Flash extends this backbone to predict the tumor immune microenvironment directly from routine H&E images, achieving 6× faster inference and 8× less GPU memory than its CNN-based predecessor. All models are released under the Apache-2.0 license.

—