Friday, July 24, 2026

1. AREX: Towards a Recursively Self-Improving Agent for Deep Research
   T2arXiv AI/CL/LG（论文）

   AREX is a recursively self-improving deep research agent framework that alternates between an inner research loop for gathering evidence and constructing a provisional answer, and an outer self-improvement loop that audits the answer constraint-wise, identifies unresolved claims, and launches targeted follow-up research. It learns an autonomous context-update tool to compress growing interaction history into a compact improvement state without relying on an external model. Trained on a 4B dense model and a 122B-A10B Mixture-of-Experts model, AREX substantially outperforms comparable-scale baselines on BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam, and other benchmarks, remaining competitive with models using substantially more activated parameters.

2. Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems
   T2arXiv AI/CL/LG（论文）

   Production AI agents often fail not due to poor reasoning but inability to manage conversation history, large prompts, tool definitions, and outputs. This paper proposes Agentic Context Management (ACM), decomposing it into five primitives: architecting, ingesting, scoping, anticipating, and compacting & consolidation. It argues that only validated compaction achieves linear token cost with preserved fidelity. Reference implementation Maximem Synap achieves 92% on LongMemEval and 93.2% on LoCoMo. The paper also discusses underexplored dimensions like latency, token efficiency, and context-rot resistance.

3. How AI Guardrails Are Impeding the Work of Offensive Cybersecurity Researchers
   T2TechCrunch：AI（RSS）

   Multiple offensive cybersecurity researchers report that AI guardrails from companies like OpenAI and Anthropic are severely disrupting their vulnerability discovery and exploit tool development. Overly restrictive content filters cause models to reject exploit code generation or return misleading information, forcing researchers to seek more open alternatives or abandon certain research directions. This highlights the adaptability flaws of current AI safety strategies in specialized security domains.

1. OpenRouter Launches Classifiers: Track AI Agent Activity and Costs by Department, Task Type
   T1OpenRouter：Announcements（RSS）

   OpenRouter introduces Classifiers, enabling users to define a taxonomy (e.g., department, task type, agent complexity) and automatically tag each generation request via a small model. Users can filter logs and group consumption (Activ) by classification, providing precise monitoring of AI agent usage patterns and cost distribution. The feature helps businesses understand where AI spend goes and optimize resource allocation.

2. AMD Launches Helios AI Rack-Scale System, Takes on Nvidia
   T2TechCrunch：AI（RSS）

   AMD announced the Helios AI rack-scale system, designed for large-scale AI training and inference, with shipments starting later this year. The system leverages AMD's next-gen GPUs and Infinity Fabric interconnect, directly targeting Nvidia's DGX and HGX series. AMD claims significant advantages in energy efficiency and total cost of ownership, potentially breaking Nvidia's dominance in AI data center hardware.

3. Runway Launches AI Model Router, Bets on Generative Media Infrastructure
   T2TechCrunch：AI（RSS）

   Runway no longer wants to be just another AI model company; it aims to become the infrastructure layer for generative media. On Thursday, the startup launched an AI model routing service designed to address the growing complexity of model selection in a crowded generative media landscape. The router automatically selects the best-fit model for each task (e.g., video generation, image editing), reducing user decision costs and improving efficiency. This marks a strategic shift from a single-model provider to a platform infrastructure player, potentially reshaping the generative media ecosystem.

4. OpenAI's New Voice Mode Arrives on ChatGPT Desktop App
   T2TechCrunch：AI（RSS）

   OpenAI has launched its advanced voice mode on the ChatGPT desktop app. The feature integrates with ChatGPT Work and Codex, enabling users to complete complex tasks and control AI agents via voice. This marks a deeper step for voice interaction in desktop assistant and automation scenarios.

5. Bluesky's AI Assistant Attie Expands into an Open Social Research Tool
   T2TechCrunch：AI（RSS）

   Users can now ask Attie questions about news, trends, and conversations on Bluesky and other apps on the AT Protocol. This marks Attie's transformation from a personal AI assistant into an open social research tool, enabling researchers and users to analyze public discussions on decentralized social networks.

6. Anthropic Updates Claude Voice Mode with Meeting Rescheduling and Email Drafting
   T2TechCrunch：AI（RSS）

   Anthropic has upgraded Claude's voice mode with more capable models, enabling users to reschedule meetings or draft emails directly via voice. This enhancement boosts Claude's practicality in office scenarios and makes the voice assistant more aligned with daily productivity needs.

7. Meta Launches 'AI Optimism' Ad with Song About Human Extinction
   T2TechCrunch：AI（RSS）

   Meta used David Bowie's song 'Five Years' in an ad promoting AI optimism, but the song is actually about humans learning they have five years left before the apocalypse. The stark contrast between the message and the music sparked online mockery and debate, highlighting cultural sensitivity issues in AI-era brand marketing.

1. Midjourney Acquires Astrology App Co-Star
   T2TechCrunch：AI（RSS）

   AI lab Midjourney has acquired the popular astrology app Co-Star for an undisclosed sum. Co-Star, known for its AI-powered personalized horoscopes, boasts millions of active users. The deal marks Midjourney's expansion from image/video generation into personal lifestyle AI applications, potentially combining generative AI with verticals like astrology and psychological analysis. Both parties will retain the Co-Star brand and team, exploring new AI-driven interactive experiences.

1. Artificial Epanorthosis: Why Large Language Models Overuse a Classical Rhetorical Figure and How to Mitigate It
   T2arXiv AI/CL/LG（论文）

   This paper finds that LLMs systematically overuse the rhetorical figure epanorthosis (self-correction), driven by promotional training data and RLHF preferences, with left-to-right generation as an amplifier. Using an Epanorthosis Index, models overuse in oratory (about twofold, near threefold in Italian) and underuse in informal Q&A. Mitigations include a one-line instruction cutting it by half to three-quarters, and an SFT adapter nearly eliminating it, with a scaling coefficient for calibration to human rates per genre. The goal is calibration, not elimination; the real risk is humans writing like machines.

2. Agentic Coding Without the Cloud: Evaluating Open-Weight LLMs on Longitudinal Data Preparation Tasks
   T2arXiv AI/CL/LG（论文）

   This paper introduces an open-source framework for evaluating open-weight LLMs on longitudinal population study data preparation tasks. The framework includes a curated ground-truth dataset, task definitions, and automated evaluation routines. Benchmarking on 20 tasks (102 variables) shows that state-of-the-art 31-35B parameter models achieve up to 87.9% average task completion, suggesting a viable path for AI-assisted data preparation in governance-restricted settings using consumer-grade hardware.

3. Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity
   T2arXiv AI/CL/LG（论文）

   Existing sufficiency-oriented methods for time-series explanation can assign high importance to spurious subsequences that support the prediction without being essential. This paper proposes TimePNS, a necessity-aware framework inspired by Pearl's counterfactual notion of necessity. It intervenes on temporal factors and measures whether the original prediction is disrupted. The two-stage design first learns an identifiable causal generative process with a sufficiency mask, then performs counterfactual interventions to derive necessity signals that refine the explanation. Experiments on synthetic and real benchmarks show that TimePNS more accurately identifies decision-critical subsequences and consistently improves sufficiency-necessity trade-offs.

4. Visual Contrastive Self-Distillation: On-Policy Self-Distillation Without External Teacher, Privileged Answers, or Visual Evidence
   T2arXiv AI/CL/LG（论文）

   This paper proposes Visual Contrastive Self-Distillation (VCSD), a simplified on-policy self-distillation method that generates asymmetric signals purely through input conditioning. An EMA teacher produces two next-token distributions under the same prompt and prefix—one conditioned on the original image and the other on a content-erased control. The token-wise log-probability difference highlights candidates specifically boosted by visual content, used to sharpen the teacher's original-image distribution for student distillation. On the ViRL39K dataset, VCSD consistently outperforms matched OPSD baselines across Qwen3-VL and Qwen3.5 models; e.g., on Qwen3-VL 2B, the seven-benchmark aggregate rises from 62.27% to 67.04%. VCSD requires no external teacher, privileged answers, visual evidence signals, reasoning traces, or additional inference cost.

5. Finite-Sample Coverage Audits for High-Recall Candidate Generation: Certification and Learning-Theoretic Design
   T2arXiv AI/CL/LG（论文）

   This paper studies the label complexity of auditing high-recall candidate generation, proving that sampling only within the candidate set cannot certify the missed relevant mass; the audit must sample the excluded pool. The authors derive a minimax lower bound in the zero-miss regime and develop an exact finite-sample toolkit using binomial and hypergeometric inversion to certify missed mass, convert it to recall via a two-pool design, simultaneously certify pre-specified families of nested candidate generators, and produce stress-test certificates against declared perturbation mechanisms. All guarantees hold under the discipline that the candidate generator (or its pre-specified family) and the audit rule are fixed before examining certification labels.

6. Barzilai-Borwein Fails Superlinear Convergence on an Open Set of Quadratics for Every Dimension n ≥ 4
   T2arXiv AI/CL/LG（论文）

   The Barzilai-Borwein (BB) method performs well in continuous optimization, yet its convergence dynamics have long been poorly understood. This paper constructs a nonempty open set (positive Lebesgue measure) of strictly convex quadratic problems and initial points in every dimension n ≥ 4, showing that the long BB1 method converges but cannot achieve root-superlinear convergence. Via a computer-assisted proof of a nonresonant, attracting seven-cycle in the projectivized BB dynamics in dimension four, it shows that gradient norms and energy norms are bounded both above and below by geometric sequences, ruling out superlinear convergence.

7. Unsupervised Consensus-Based Anomaly Detection Reveals Spatiotemporal Malaria Anomalies in Ghana
   T2arXiv AI/CL/LG（论文）

   This study applies a consensus anomaly detection framework to monthly malaria surveillance data from Ghana (2014-2023), identifying highly structured spatiotemporal anomaly patterns. Ashanti and Northern Regions are recurrent hotspots, with persistent anomalies in Tamale, Kumasi, and Accra. Key finding: spatial distinction between anomaly burden (cumulative cases during anomalies) and anomaly frequency (persistence of unusual behavior). Tamale has highest burden during anomalies, but highest anomaly rates cluster in Ashanti districts. Anomalous months differ significantly from normal months (Cohen's d = 3.252). The framework separates high-prevalence areas from behaviorally unusual areas, strengthening surveillance and targeted control.

8. Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context
   T2arXiv AI/CL/LG（论文）

   This paper addresses the draft model cost issue in speculative decoding under million-token contexts by proposing Windowed-MTP. It applies a StreamingLLM-style sliding window plus attention sink only to the draft's attention, bounding the KV working set to a constant and eliminating the full-context draft-KV read tax while preserving full-attention verification in the target model. Experiments on Qwen GDN-MoE 35B/122B and Mamba2-hybrid 120B in single-GPU SGLang show a 28%-44% reduction in per-decode-step cost with no loss in output distribution. Additionally, unread draft KV cache (~7.7%-11%) is reclaimed via a compact ring buffer.

—