Thursday, July 2, 2026

1. Claude Code v2.1.198 Released: Chrome Integration and Background Agent Notifications
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code v2.1.198 highlights: Claude in Chrome is now generally available; background agent notifications added (hooks agent_needs_input / agent_completed); new /dataviz skill for chart/dashboard design with a runnable color-palette validator; Gateway adds Anthropic AWS as an upstream provider. Multiple fixes include network retry on transient errors, background tasks stuck on 'Running', markdown table overflow, macOS background agent local network routing, .claude/rules symlink loading, and more. Improvements to focus mode, syntax highlighting, and keyboard shortcuts.

2. QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling
   T2arXiv AI/CL/LG（论文）

   Scaling inference compute by generating many independent attempts per problem is a reliable but costly lever for improving language models, leading to redundant solutions. QuasiMoTTo uses quasi-Monte Carlo to generate correlated but exact samples as a drop-in replacement for i.i.d. samples, improving sample efficiency without sacrificing parallelism. Across four reasoning benchmarks, QuasiMoTTo matches i.i.d. pass@k accuracy with 25-47% fewer samples, and in policy-gradient RL (GRPO) it matches i.i.d. performance with 50% fewer training steps. By spreading samples more evenly over the output space, it reduces redundancy and strengthens the learning signal per batch.

3. AutoMem: Automated Learning of Memory as a Cognitive Skill
   T2arXiv AI/CL/LG（论文）

   AutoMem proposes treating memory management as a trainable skill for LLMs, automating two loops: one uses a strong LLM to review full trajectories and iteratively revise memory structures, the other extracts good memory decisions from many episodes as training signals. In three long-horizon games (Crafter, MiniHack, NetHack), optimizing memory alone improves baseline performance ~2-4x, making a 32B open-weight model competitive with Claude Opus 4.5 and Gemini 3.1 Pro Thinking. Results show memory management is an independently learnable skill with high leverage for long-horizon tasks.

4. Right in the Right Way: LM Training with Verifiable Rewards and Human Demonstrations
   T2arXiv AI/CL/LG（论文）

   This paper proposes an adversarial generator-discriminator framework that augments verifiable rewards with a learned signal from human demonstrations. The generator is trained via RL to maximize both task accuracy and an adversarial reward from the discriminator, which learns to distinguish human-written outputs from model-generated ones. Across bug fixing, story generation, and a reward hacking benchmark, the method consistently improves non-verifiable properties while preserving RLVR accuracy gains and nearly eliminating reward hacking.

5. Diffusion-GR2: Diffusion Generative Reasoning Re-ranker
   T2arXiv AI/CL/LG（论文）

   Diffusion-GR2 converts an autoregressive reasoning re-ranker into a diffusion model, achieving 2.4–3.5× decode throughput gain via parallel denoising, while closing accuracy gaps through conversion fine-tuning, on-policy distillation, and reinforcement learning, reaching near-parity on Amazon Beauty.

6. Sequentially-Controlled Interactive Multi-Particle Flow-Maps for Online Feedback-Driven Search
   T2arXiv AI/CL/LG（论文）

   This paper proposes the IMPFM framework, which uses multi-particle interactive flow-maps for sample-efficient online feedback-driven search, addressing broad exploration when preferences are unknown a priori. It features a posterior sample sharing mechanism and an exploration-exploitation reweighting strategy, with theoretical proof that it steers the multi-particle system toward a KL-tilted target distribution, preventing mode collapse. Extensive empirical evaluations show IMPFM significantly outperforms existing baselines across diverse search and alignment tasks.

7. Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains
   T2arXiv AI/CL/LG（论文）

   Large Language Model (LLM) agent skills, while packaging reusable operational knowledge, suffer from opaque identity, version, and provenance information as they proliferate, causing duplicate dependencies and inconsistent installations. This paper introduces Agent Skill Supply Chains (ASSCs) to characterize mixed skill-package-service dependency graphs, and designs SkillDepAnalyzer—inspired by Software Bill of Materials (SBOMs)—to capture natural-language dependency evidence model skills as dependency-bearing artifacts. On the SKILL-DEP benchmark, SkillDepAnalyzer accurately recovers skill metadata and dependency graphs, outperforming LLM-based baselines and package-centric SBOM tools. Applied to over 1.43 million skills, the analysis reveals four structural patterns, hidden security signals in dependencies, and identifies known malicious skills persisting in ASSCs. Findings recommend typed dependency manifests, first-class dependency-cluster management, and lockfile-like records for skill developers.

8. Autonomous Scientific Discovery via Iterative Meta-Reflection: Introducing DiscoPER
   T2arXiv AI/CL/LG（论文）

   DiscoPER is an autonomous LLM-powered framework that conducts open-ended research by dynamically generating and executing code to explore datasets without pre-specified objectives. It features a second-order reasoning mechanism that periodically analyzes its accumulated discoveries to identify structural patterns and epistemic gaps, redirecting hypothesis exploration toward uncharted regions. Evaluated on the iNatDisco multimodal ecological benchmark, it recovers 8 of 9 known patterns with a 72.7% hypothesis support rate, outperforming classical causal discovery and LLM-guided baselines. This work demonstrates the potential of AI for open-ended scientific discovery.

1. Theoria: Rewrite-Acceptability Verification over Informal Reasoning States
   T2arXiv AI/CL/LG（论文）

   Theoria is a novel verification architecture that rewrites an AI system's answer into a sequence of typed state transitions, each licensed by an explicit justification (citation, computation, or problem-given fact), enabling independent auditing of every step. On HLE-Verified Gold, Theoria certifies 105 problems at 91.4% strict precision, producing human-readable proof traces. Against adversarial poisoned proofs, it catches 94.7% vs. 83.2% for holistic judges, with advantages concentrated in hidden premises (90.6% vs. 62.5%) and fabricated citations (100% vs. 90%). This architecture offers an auditable and reproducible path to trustworthy LLM outputs.

2. Are Performance-Optimization Benchmarks Reliably Measuring Coding Agents?
   T2arXiv AI/CL/LG（论文）

   This study systematically audits the reliability of three performance-optimization benchmarks: GSO, SWE-Perf, and SWE-fficiency. By replaying official reference patches for 740 tasks, it finds that most tasks fail to satisfy original validity rules across cross-machine replays, with SWE-Perf being especially fragile. Additionally, public submission rankings are highly dependent on scoring rules, and most tasks are already solved by at least one submission. The research exposes performance gaps hidden by aggregate rankings.

3. GPU-Parallel Linearization Error Bounds for Real-Time Robust Optimal Control of Nonlinear and Neural Network Dynamics
   T2arXiv AI/CL/LG（论文）

   This paper proposes tight, differentiable GPU-parallel linearization error bounds (LEBs) for LTV approximations of nonlinear and neural network dynamics. For analytic dynamics, path-based Hessian bounds are introduced; for NN dynamics, certified LEBs are derived using affine relaxations from NN verifiers and local Jacobian corrections. The GPU-parallel system-level synthesis LTV robust control solver is extended to handle right-invertible disturbance matrices and non-zero-centered disturbance sets, enabling online optimization of robust feedback policies accounting for linearization error. On nonlinear and NN dynamics up to 168 state dimensions, the method achieves 67Hz GPU computation, reducing solve times and conservativeness while preserving formal guarantees and real-time performance.

4. Disentangling Speaker and Language Effects in Cross-Lingual Speaker Verification for Iberian Languages
   T2arXiv AI/CL/LG（论文）

   Cross-lingual speaker verification (SV) systems typically degrade when enrollment and test utterances are in different languages, but standard protocols confound language mismatch with inter-speaker variability. This work introduces a bilingual same-speaker evaluation set for five Iberian languages, enabling analysis under constant speaker identity. Using a HuBERT-based SV system and the Cross-Lingual Transfer Matrix (CLTM), results show speaker variability accounts for part of the degradation, but language mismatch remains the primary driver. These findings provide a more precise characterization of language dependence in cross-lingual SV.

5. Indian Tech Tycoon Bets $30M of His Own Money to Build an AI-Powered Alternative to Microsoft Office
   T2TechCrunch：AI（RSS）

   Indian tech entrepreneur Bhavin Turakhia has announced a $30 million personal investment in Neo, his fifth venture, aiming to build an AI-powered alternative to Microsoft Office and Google Apps. The company focuses on AI-driven document collaboration and productivity enhancement for enterprises.

6. Ashton Kutcher Leaves Sound Ventures, Launches New VC Firm with Morgan Beller Focusing on AI Infrastructure and Energy
   T2TechCrunch：AI（RSS）

   Ashton Kutcher is leaving Sound Ventures, the firm he co-founded, to launch a new VC firm with former partner Morgan Beller. While Sound Ventures was known for concentrated, high-conviction bets on leading AI labs, the new fund will focus on the underlying infrastructure and energy powering those companies, signaling a strategic pivot.

7. Cloudflare Sets Deadline for AI Firms to Distinguish Crawlers, Paving Way for Publisher Compensation
   T2TechCrunch：AI（RSS）

   Cloudflare issues a policy requiring AI companies to separate web crawlers for search from those for AI training and agents by September 15, or face default blocking on many publisher sites. The move enables publishers to identify and charge for content used by AI, addressing concerns over uncompensated scraping.

8. A Two-Step Method for Optimal Resource Utilization in Autonomous Laboratories
   T2arXiv AI/CL/LG（论文）

   In autonomous laboratories, AI agents propose batches of experiments, but resource planning faces hardware constraints. This paper presents a two-step method for a metal-organic framework synthesis platform: first, constraint programming finds optimal schedules minimizing total time; second, a task status dependency system ensures robust execution. The approach improves resource utilization under heterogeneous multi-instrument conditions.

—

1. Neural Certificate Pricing for Combinatorial Optimization Problems
   T2arXiv AI/CL/LG（论文）

   Combinatorial optimization (CO) problems are difficult due to exponential search, but structural feasibility of a solution can be verified in polynomial time once supplied. This paper introduces Neural Certificate Pricing (NCP), leveraging this asymmetry under unsupervised learning. A neural network predicts certificate-level dual prices, and a structured recovery layer constructs the induced primal marginal. Under certificate-consistency, the recovered marginal is globally feasible; first-order errors in price induce only second-order loss in objective value. Across three CO problem classes, NCP significantly outperforms or matches state-of-the-art neural baselines with much less computation time, showing stronger out-of-distribution generalization.

—