Thursday, August 20, 2026

1. Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles
   T2arXiv AI/CL/LG（论文）

   This paper proposes viewing the leaf value of each tree in a gradient-boosted ensemble as a coordinate rather than an intermediate result, mapping every instance to a point in R^M where the model acts linearly. This change of perspective makes contrastive explanation exact: the difference between two instances is nonzero only on coordinates where they do not share a leaf, and each such coordinate is traceable to a real split. The recourse method built on this representation reconstructs the model's own decision to 6.2×10^-15 across five tabular datasets under repeated cross-validation. On credit datasets it is Pareto-non-dominated on effort against realism. When restricted to changes the subject could actually make (excluding age or settled delinquency), it retains 58% validity versus 41% for the strongest baseline, a distinction standard evaluation misses.

2. Continuous-Time Reinforcement Learning for Controlled Hawkes Jump-Diffusions
   T2arXiv AI/CL/LG（论文）

   This study addresses stochastic control of multivariate Hawkes-driven stochastic differential equations using machine learning algorithms in a non-Markovian setting. Due to the path dependence of Hawkes intensity memory, the problem does not fit classical stochastic control theory. The authors develop a finite-dimensional Markovianization procedure approximating multivariate Hawkes processes with mixtures of exponential kernels and prove convergence. They then formulate continuous-time deterministic policy gradient learning, called Hawkes-CT DDPG, which solves the non-Markovian optimization model-free by observing only event times, SDE realizations, and decay filters, without knowing Hawkes kernel coefficients. Comparisons against discrete-time reinforcement learning under exponential, Erlang, and power-law kernels demonstrate the effectiveness of the proposed method.

3. Grouping the Stochastic Machine: Precision, Not Capability, as the Frontier Metric for AI Systems
   T2arXiv AI/CL/LG（论文）

   Frontier language models have saturated on capability: their average output lands on the target. This paper argues that what now differentiates systems is not capability but precision — how tightly outputs cluster around the target across repeated identical requests. Using a marksman's distinction, the author defines a grouping metric that is cheaply measurable by running a fixed suite of deterministically scored tasks many times at fixed temperature and computing per-task consistency. This metric separates consistent, off-centre failures (correctable by a sight adjustment) from scattered failures (requiring a different model or sampling). A first real run showed a single rule closing a measured gap from 0/5 to 5/5, while rule-authored tasks found no value—proving that a discipline's worth must be measured on real work, not derived from its own rulebook.

1. Mistral AI Launches Agentic Search for Deep Document Retrieval
   T1Mistral AI：News（网页）

   Mistral AI has recently unveiled Agentic Search, a novel retrieval layer engineered to help AI systems navigate, read, and verify information across even the most intricate documents. Unlike conventional keyword-based search, this agentic approach enables the AI to actively explore document structures, cross-reference related content, and validate data points, reducing misinformation and hallucinations. The new tool is poised to benefit sectors like legal, finance, and academic research, where precision in handling complex documents is critical. By integrating this capability, Mistral AI aims to enhance the reliability and depth of AI-driven insights. This launch also signals a strategic shift for the company, moving beyond open-source model development to offer comprehensive AI infrastructure solutions. For developers and enterprises, Agentic Search presents an accessible yet powerful tool to build more robust applications, potentially setting a new standard for intelligent information retrieval. The announcement reflects Mistral's commitment to open and advanced AI technologies.

2. OpenAI Reaffirms Zero Data Retention for Frontier Models, Previews Private Safety Processing
   T1OpenAI：News（RSS）

   OpenAI has reaffirmed its commitment to Zero Data Retention for eligible API customers using frontier models, ensuring that prompts and responses are not stored by OpenAI. In addition, the company previewed its new Private Safety Processing capability, which allows advanced safety reviews to be performed without exposing raw data, preserving privacy while maintaining robust oversight. This development comes as enterprises in regulated industries increasingly demand both strict data governance and access to cutting-edge AI models. By combining zero-storage guarantees with privacy-preserving safety checks, OpenAI aims to bridge the gap between compliance and model safety. The preview suggests that future deployments of frontier models may adopt these privacy-preserving techniques more broadly, potentially setting a new standard for secure AI interaction. The company has not yet disclosed a general availability timeline, but the announcement signals a strategic focus on enterprise trust and data protection.

3. OpenAI Offers Zero Data Retention for Frontier Models and Previews Private Safety Processing
   T1OpenAI：News（RSS）

   OpenAI has announced Zero Data Retention (ZDR) for eligible API customers using frontier models, ensuring prompts and responses are not stored. The company also previewed Private Safety Processing (PSP), a mechanism that enables advanced AI safety checks without exposing private data. This addresses the conflict between enterprise data privacy and AI safety. The move is aimed at regulated industries like finance and healthcare, where compliance is critical. With ZDR, sensitive information will not be used for training or retained by OpenAI. PSP demonstrates a path toward privacy-preserving safety oversight. This could make privacy a key differentiator among AI providers. Enterprises will need to watch for further details on eligibility and verification. Overall, it marks an important step for trustworthy AI adoption.

4. Claude Code v2.1.236 Release: Adds Default Model Variable and Fixes Multiple Stability Issues
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code released v2.1.236. A new ANTHROPIC_DEFAULT_MODEL environment variable sets the default model for new sessions, while a /model pick still overrides it and persists across restarts. SendMessage now supports notify_when_idle, allowing another session to send a one-shot notice when it next goes idle without polling. On macOS sandbox, wildcard read-deny rules take precedence inside allowed read regions, cover directory contents, and cannot be bypassed by renaming. The release also fixes issues with clipboard, fullscreen rendering, tmux titles, and improves auto mode classification and startup performance.

5. OpenAI: How ChatGPT Work and Codex Help Stampli Move Ideas to Market Faster
   T1OpenAI：News（RSS）

   OpenAI published a customer story about Stampli, a finance automation platform, which used Codex and ChatGPT Work to compress weeks of launch production into days despite limited design resources and a fixed deadline. The case highlights the potential of generative AI tools to shorten development cycles and improve delivery efficiency in real business settings.

6. Claude Code v2.1.237 Released: Fixes Gateway Prompt Caching, Adds 'Concise' Output Style
   T1.5Claude Code：GitHub Releases（RSS）

   Anthropic released Claude Code v2.1.237, focusing on fixing prompt caching for sessions using an LLM gateway or custom base URL, improving cache hit efficiency and response stability. It also introduces a built-in 'Concise' output style where Claude leads with results and skips preamble and narration, reducing verbosity for developers who want efficient command execution. This update targets developer experience, strengthening Claude Code's competitiveness among AI coding tools.

7. Google launches new AI study tools across Search and Gemini
   T2TechCrunch：AI（RSS）

   Google today announced new AI study tools across Search and Gemini, designed to help students learn and prepare for exams more effectively. Search introduces "study cards" and step-by-step problem solving, offering visual explanations of concepts and guided solutions. Gemini gains a personalized learning assistant that can generate quizzes, summarize notes, and track study progress from course materials. The rollout is seen as Google's latest move to compete with OpenAI in the education AI space. The features will be rolled out to users over the coming weeks.

8. OneCLI: An Open-Source Sandboxed Agent Harness for Teams
   T2Hacker News 热门（Algolia · 中文化）

   OneCLI, an open-source sandboxed agent harness from Y Combinator's S26 batch, aims to provide a secure personal AI agent for every employee in a team. Unlike generic agent frameworks, OneCLI emphasizes sandboxing and permission controls, allowing organizations to deploy AI agents without exposing sensitive data or critical systems. It integrates with existing development workflows, offering an isolated execution environment, auditability, and governance features. The project was launched on Hacker News and has attracted attention from engineers and enterprises seeking safe, team-wide AI adoption. By open-sourcing the harness, OneCLI enables customization and self-hosting, which is particularly appealing for security-conscious companies.

1. Conceptual Integrity and Lines of Code: Software Engineering in the Age of AI Coding Agents
   T2Simon Willison 博客（Atom）

   In a recent Talking Postgres podcast episode, Simon Willison and Claire Giordano discussed how AI is changing software development. Willison argues that lines of code can remain a meaningful productivity metric when using coding agents: while a human might produce 50–200 production-ready lines per day, agents can push that to a thousand, provided quality is maintained. However, the new bottleneck is cognitive capacity, not typing speed. He invokes “conceptual integrity” from The Mythical Man-Month, warning that AI makes adding features so cheap that software can grow into a “Winchester Mystery House” of incoherent additions. Discipline and senior engineering judgment become even more essential.

2. OpenAI unveils new customer privacy protections to outdo Anthropic in enterprise market
   T2TechCrunch：AI（RSS）

   On August 19, 2026, OpenAI announced a new set of customer privacy protections aimed squarely at enterprise clients, intensifying its rivalry with Anthropic. The measures include advanced data isolation, stronger commitments to not train models on customer data by default, and more transparent compliance and audit capabilities. This move directly challenges Anthropic's earlier privacy-certified offerings, as both AI labs compete to become the default choice for regulated industries. Enterprises have become increasingly wary of how their proprietary data is used in AI training, and the race reflects a broader industry shift toward privacy as a key differentiator. OpenAI's announcement is designed to reassure large enterprises, especially in finance and healthcare, that their sensitive data remains protected, while also signaling that the competition over AI trust in the enterprise sector has entered a new phase. Analysts say the outcome could shape how AI providers structure data governance and client agreements in the coming years.

3. AI Adoption Isn't Winning Hearts: Why Silicon Valley Is Surprised
   T2TechCrunch：AI（RSS）

   A new TechCrunch report suggests that as AI becomes harder to avoid, consumers are growing more wary of the technology. Silicon Valley once believed that widespread deployment would naturally win public approval, but it now faces a clear gap between adoption and acceptance. The report argues that AI companies must address growing demands for transparency, safety, and user control while pursuing scale — otherwise they risk stronger social backlash and regulatory pressure. This marks a critical moment for the industry to rethink how trust is built into AI products rather than treating adoption as a proxy for approval.

1. ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning
   T2arXiv AI/CL/LG（论文）

   ADEPT is a large-scale reinforcement learning framework designed to accelerate sim-to-real transferable dexterity across high degree-of-freedom robot embodiments via pre-training and post-training. It first pre-trains a dexterous policy on a generic object reposing task, then uses this pretrained behavior as a prior for post-training downstream policies, avoiding learning from scratch. The framework introduces a stable post-training recipe combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates, along with a joint-space Geometric Fabric. It demonstrates zero-shot sim-to-real transfer on a 23-DoF Kuka-Allegro and a 29-DoF Flexiv-Sharpa, solving long-horizon tasks at human-level speed.

2. SCORE: Subject Coordinate Recovery for Label-Free Cross-Subject EEG-to-Image Retrieval
   T2arXiv AI/CL/LG（论文）

   EEG-to-image retrieval holds great promise for neural communication, but performance drops sharply for new users without labeled calibration. This paper analyzes EEG features and finds that different subjects preserve similar concept relationships but express them along different coordinate directions, leading to the Subject Coordinate Recovery (SCORE) framework. During training, SCORE aligns source-subject EEG with a common image space and simulates unseen-subject recovery via source-only episodes. At deployment, with frozen encoders, it selects reliable EEG-image landmarks through hubness-corrected matching and estimates an orthogonal transformation to recover target coordinates without source data or target labels. On THINGS-EEG2 and Alljoined-1.6M, SCORE outperforms the strongest baselines by significant margins in 200-way retrieval, bringing practical cross-user visual decoding closer to deployment.

3. Beyond Trial Averaging: Anchoring Neural and Visual Representations for Few-Repetition Brain-to-Image Retrieval
   T2arXiv AI/CL/LG（论文）

   Brain-to-image retrieval typically requires averaging up to 80 neural trials per image to suppress noise; when only one or a few repetitions are available, retrieval accuracy drops sharply. This paper discovers a non-transitive alignment pattern: low-repetition query signals and image representations each align with the high-repetition center, but not directly with each other, indicating that query noise is only part of the problem and gallery placement also matters. The authors propose NEAR, a neural-anchor-based retrieval framework that treats the high-repetition center as an anchor and approaches it from both sides: a denoiser pulls the noisy query toward the true anchor, and a small network predicts each candidate's pseudo-anchor from its image and pulls the image toward it. Across four datasets spanning EEG, MEG, and fMRI, NEAR consistently improves retrieval in the few-repetition regime. On THINGS-EEG2, it improves 200-way Top-1 accuracy by 5.7 and 9.3 percentage points when averaging one and four repetitions, respectively. By anchoring neural and visual representations, NEAR reduces reliance on repeated acquisition and brings neural retrieval closer to real-world deployment.

4. What is Missing from AI Post-Training: An Empirical Analysis
   T2arXiv AI/CL/LG（论文）

   This paper investigates the ability of LLM agents to post-train LLMs end-to-end, arguing that this process conflates execution-level capability with strategy-level capability. By analyzing a large corpus of public post-training trajectories, the authors find that the agent's training strategy is locked in at the very beginning, with the remaining budget spent on local adjustments. Experiments show that an experience-driven scaffold improves execution, human guidance can redirect the initial strategy, and additional inference compute helps only on simpler tasks. The conclusion is that agents lack a mechanism for spontaneously reevaluating their strategy during execution.

5. Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation
   T2arXiv AI/CL/LG（论文）

   Multi-teacher on-policy distillation (M-OPD) consolidates domain-specialized RL experts into a single generalist student via dense token-level rewards, yet its optimization dynamics remain poorly understood. This paper builds a controlled benchmark on SmolLM3-3B-Base and finds that standard M-OPD captures only 35.6% of the headroom relative to an oracle ensemble. The failure stems from misallocation of the token-level optimization budget, not gradient conflict, driven by sequence-length disparities, non-uniform learning rates, and multi-step reward staleness. The proposed Open-MOPD framework introduces token-share balancing, gap-aware dynamic budget allocation, and student reward refresh, lifting headroom recovery to 83.4%. All post-training recipes, trajectories, and evaluation suites are open-sourced.

6. Beyond the Transcript: Detecting Covert Coordination in Latent Multi-Agent Communication
   T2arXiv AI/CL/LG（论文）

   Language-model agents can communicate through continuous hidden states invisible in public transcripts, enabling covert harmful coordination. This paper introduces Verifiable Latent Alignments (VLA), an activation-aware framework that monitors and steers these private channels using representation anomaly detection, counterfactual action-distribution influence, and sparse-autoencoder interpretation. On a controlled auction benchmark, monitoring achieves a mean AUROC of 0.993 for homogeneous agents and 0.854 for heterogeneous pairs; full white-box steering recovers bid distributions exactly and reduces collusive low-bid behavior by 47.3 percentage points.

7. Comment-level Topic Drift Analysis in the Reddit Corpus
   T2arXiv AI/CL/LG（论文）

   This paper presents a novel embedding-based dynamic topic modeling approach to detect and quantify topic drift at the comment level in a massive corpus. Using pretrained language models to generate contextualized embeddings for short text, the authors analyze 12.7 billion Reddit comments from 2006 to 2022. Unsupervised methods identify dynamically evolving topic clusters, while a null model comparison test filters out spurious dynamics. Key findings show that politically and socially contentious topics exhibit significant directional drift in embedding space, with inter-topic distances changing systematically beyond null expectations, whereas music and sports remain comparatively stable. The work offers a scalable methodology for studying semantic and discourse evolution.

8. Enhancing EBSD Throughput of Battery Electrode Materials Using Super-Resolution Generative Adversarial Networks
   T2arXiv AI/CL/LG（论文）

   This study proposes a super-resolution framework based on generative adversarial networks (SRGAN) to enhance the throughput of electron backscatter diffraction (EBSD) analysis for Li-ion battery electrode materials. The model was trained on EBSD data of NMC cathode particles to computationally upscale low-resolution datasets by factors of 2x to 12x, and its performance was compared with classical interpolation methods. Results show that SRGAN systematically outperforms classical methods in preserving small grains and maintaining realistic grain boundaries. At 5x upscaling, corresponding to a 25x speed-up in acquisition time or a 25x larger field of view, the relative errors for grain area-equivalent diameter, grain maximum sphere-inscribed diameter, and grain boundary length were +5.7%, +8.2%, and -14.6%, respectively. This methodology significantly enhances EBSD efficiency, enabling high-throughput characterization for materials research and industrial process development.

—