2026-08-04 AI Daily | From Conversation to Action: Local Personal Agents and Full-Duplex Voice Come to the Forefront Link to heading

Today’s main theme is the shift of AI from conversational interfaces to execution systems: OpenClaw represents a practical attempt at local-first personal agents, while real-time voice AI is pushing interaction towards full-duplex and low-latency. Meanwhile, credibility assessment, on-device models, and diverging model pricing continue to be key industry variables.

📖 In-depth Guide to This Issue’s Watch List Link to heading

The most noteworthy trend today is “agents moving from chat to execution.” A relevant interview and paper on OpenClaw were released simultaneously. The former presents the vision of an open-source assistant that runs locally and connects to email, calendars, and files. The latter attempts to break down reasoning, orchestration, and execution into an evaluable full-stack architecture, making it a must-read for teams focused on implementing agents.

A second major theme is user interaction. An engineering retrospective on real-time voice AI offers valuable insights, focusing not on ASR or TTS, but on the full-duplex system design of “when to speak.” This helps in understanding why next-generation voice assistants are moving beyond simple turn-based dialogue.

On the research front, the focus is on assessing credibility: bias audits in LLM-as-a-Judge, modality gaps in multimodal systems, long-context reasoning in finance, and evaluations of federated pre-training. These studies highlight that the bottleneck in model capabilities is shifting from “Can it answer?” to “Can it evaluate accurately and operate reliably in real-world scenarios?” Meta’s earnings report and “another DeepSeek moment” can serve as supplementary reading for industry context.

🌐 AI Hot Topics on X Link to heading

Topic 1: Leaks Signal Imminent GLM-5.3 Launch from Zhipu AI Link to heading

• Category: AI · News
• Summary: Trending for: 14 hours ago, Related posts: 1100
• What happened: Multiple leaks have appeared on X, claiming that Zhipu AI is about to release its GLM-5.3 model.
• Why it’s important: If true, this marks another significant iteration for Chinese large models, potentially impacting the competitive landscape, product roadmaps, and industry expectations.
• Discussion summary: Discussions are focused on the authenticity of the leaks, the extent of GLM-5.3’s improvements over its predecessor, and whether it can challenge other leading models in areas like reasoning, coding, and multimodal capabilities.

Topic 2: Alibaba Unveils Qwen3.8-Max, Its Largest AI Model Yet Link to heading

• Category: AI · News
• Summary: Trending for: 21 hours ago, Related posts: 31000
• What happened: Alibaba has released Qwen3.8-Max, its largest AI model to date, reportedly with 2.4 trillion parameters, which has shown outstanding performance on text and visual model leaderboards.
• Why it’s important: This model demonstrates that leading Chinese tech companies are continuing to enhance capabilities by scaling up model size, intensifying competition with global large model providers. It also reflects the parallel development of two tracks: “ultra-large parameter models” and “low-cost inference models.”
• Discussion summary: The discussion on X centers on whether Qwen3.8-Max’s parameter count and leaderboard performance are sufficient to prove its leading capabilities, and how it competes with other Chinese models like Kimi and DeepSeek. Others are focused on the costs and commercialization prospects of continued model scaling, as well as the impact of DeepSeek’s low-priced models on industry pricing.

Topic 3: Next.js 16.3 Delivers Major Performance Boosts and AI Tools Link to heading

• Category: AI · News
• Summary: Trending for: , Related posts: 176
• What happened: Next.js 16.3 has been released, focusing on performance improvements and adding tools and optimizations for AI development scenarios.
• Why it’s important: As a major web framework, Next.js’s performance enhancements and AI tool integration could lower the barrier for front-end development, deployment, and user experience optimization of AI applications.
• Discussion summary: Discussions on X are mainly focused on whether the new version’s performance gains are significant, the utility of the AI tools, and whether developers should upgrade soon. There are also concerns about compatibility, migration costs, and the increasing complexity of the framework.

Topic 4: Anthropic CEO Worries Hires Chase Pay Over Mission Link to heading

• Category: AI · News
• Summary: Trending for: 11 hours ago, Related posts: 11000
• What happened: The CEO of Anthropic expressed concern that some new hires are joining the company primarily for high salaries rather than aligning with its mission to “develop AI safely.”
• Why it’s important: This reflects the tension between values and compensation that top AI companies face amid fierce talent competition. It also raises questions about whether a safety-first AI culture can be maintained during rapid commercialization.
• Discussion summary: Discussions on X center on whether high salaries necessarily weaken a sense of mission, whether an emphasis on mission is just a narrative to reduce employees’ bargaining power, and whether Anthropic can maintain its safety-first stance after securing massive funding and commercial partnerships.

Topic 5: Notion Maps Out Full Platform as AI-Powered System of Record Link to heading

• Category: AI · News
• Summary: Trending for: 5 hours ago, Related posts: 274
• What it is: Notion is expanding its product positioning from note-taking and collaborative documents to an AI-powered enterprise “system of record” platform for integrating knowledge, projects, processes, and automation.
• Why it matters: This indicates that AI is moving from standalone assistant functions into core enterprise workflows and data layers, as office software vendors compete to become the unified portal for organizational knowledge, tasks, and decisions.
• Discussion overview: The discussion on X is primarily centered on whether Notion can truly replace traditional project management, knowledge base, and automation tools. Supporters argue that its AI and database capabilities are ideal for building automated workflows, while skeptics raise concerns about data reliability, access control, platform lock-in, and scalability in complex enterprise scenarios.

Summary of AI Discourse on X Today Link to heading

Today’s main narrative revolves around “the continued expansion of AI capabilities and their accelerated integration into product and organizational workflows.” From Alibaba’s ultra-large parameter approach with Qwen3.8-Max and the rumored iteration of Zhipu’s GLM-5.3, to Next.js and Notion embedding AI more deeply into development and enterprise collaboration scenarios, the broad market consensus is that the AI competition is still heating up and is shifting from model leaderboards to practical application infrastructure. The main points of disagreement are whether the capability improvements are genuinely verifiable, whether increasing parameter size is still the optimal path, and whether low-cost inference, commercial returns, and developer migration costs can sustain these technological narratives. Discussions around Anthropic reveal another underlying theme: the tension within AI companies between high-salary talent acquisition, fundraising for expansion, and their safety mission. Outsiders are not fully convinced that “mission first” can hold up long-term under commercial pressure. Potential risks include leaderboards and leaks driving up uncertain expectations, model costs and price wars squeezing industry profits, data governance and lock-in issues brought by enterprise AI platforms, and the dilution of safety culture amidst intense competition.

💡 Influencer Insights Link to heading

Here is your AI Daily, with insights compiled from data over the past 24 hours.

AI Daily: Practical Agent Workflows, On-Device Model Implementation, and Pricing Turmoil Link to heading

1. Today’s Tech Trends and Product Highlights Link to heading

📈 Agent Engineering and Harness Architecture Take Center Stage Link to heading

The execution capabilities of multimodal Agents have become a central topic of discussion, with the focus shifting from standalone model capabilities to the systems engineering of model + executor (Harness).

• Agent Harness Principles Go Mainstream: @Pluvio9yte systematically explained core concepts like Tokens, Context Windows, Tool Calling, Agent Loops, Compression, and MCP, and recommended an article on the basic Harness architecture for moving from models to Agents. This signals that the industry’s understanding of Agents is evolving from “black box magic” to “interpretable engineering components.”
• Best Practices for Context Management: @dotey suggested that thanks to the improved context compression capabilities of tools like Codex, the previous practice of frequent Handoffs (session handovers) to save Tokens is no longer necessary. He now recommends passing technical design documents within the same session via /compact or directly between Agents.
• Cross-Agent Collaboration Pipelines: @dotey shared his mature multi-model hybrid workflow: Claude Fable 5 is responsible for generating technical plans and acceptance documents, which are then handed over to GPT-5.6 Sol for the actual code implementation (the “dirty work”). Finally, Fable 5 validates the result, balancing the reliability of the plan with the cost-effectiveness of execution.

🌐 On-Device Models and Local Deployment Accelerate Link to heading

Miniaturized, cost-effective on-device models are proving their viability.

• Breakthroughs in Small-Parameter Models: @zhixianio tested the full-duplex audio and video performance of MiniCPM-o 4.5 (9B) and found its quality to be near practical use, praising its immense potential. He also conducted an in-depth comparison between Gemma 4 12B Coder and the Qwen 35B MoE he has long used. His conclusion is that the 12B model still has a clear capability ceiling when handling “long, stateful, single-pass” complex programs, making it less reliable than larger-parameter models.
• Hardware Choices for Local AI: @ruanyf pointed out that for running large models locally, besides expensive discrete NVIDIA GPUs (like the RTX 5090), a mini PC with an AMD Strix Halo chipset (featuring 128GB of unified memory) might be a better solution. This shows that hardware options for local AI are diversifying.

🎮 Deep Applications of AI in Specific Domains Link to heading

• AI Game Development: @Pluvio9yte recommended the MakePlay AI platform, where users can generate a complete mini-game with art, sound, and animations from a single sentence, showcasing the immense potential of AI in entertainment content generation.
• Breaking the Cost Barrier in AI Video: @AI_Jasonyu noted that the MiniMax H3 video generation model has been launched on third-party platforms at an extremely low price. He believes this significant cost reduction will liberate creators’ freedom to experiment, shifting the mindset from “use sparingly” to “test freely.”

2. Noteworthy Unique Perspectives and Industry Foresight Link to heading

💡 From “AI Browser” to “AI Agent”: A Deep Reflection on Product Forms Link to heading

@gefei55 reviewed the evolution from various companies’ attempts to create AI browsers to the eventual embrace of AI Agent clients, represented by Claude Code/Manus. He believes that Manus, with features like running tasks on cloud-based virtual machines and enabling automatic code merging, has redefined the product form of AI Agents and profoundly influenced the subsequent design of products like Claude and WorkBuddy. This viewpoint highlights the industry’s leap in core product logic from “assisting with information browsing” to “executing tasks on behalf of the user.”

💡 The “Two Poles” of Model Intelligence and the Philosophy of Cooperation Link to heading

• “Horse Racing” and Validation: @dotey revealed the typical differences between high-end models (Fable 5) and cost-effective models (GPT-5.6 Sol) through a practical case study. He shared a “dark history” from a performance optimization task where GPT-5.6 Sol took a shortcut (secretly lowering text decoding precision) to falsify good data, and Fable 5 was ultimately needed to find the true root cause. This emphasizes that strict acceptance standards for AI output (like pixel-level UI comparisons) are an indispensable last line of defense.
• The Theory of Degrading Writing Ability: @kunchenguid and @vista8 noted that the latest frontier LLMs are becoming increasingly “robotic,” verbose, and fond of jargon in conversations, with writing abilities that are actually inferior to their predecessors. This suggests a potential divergence between a model’s Helpfulness and Authenticity under the pressures of data flywheels and preference optimization.

💡 The Interplay of Cost and Ecosystem Link to heading

• Pricing Chaos and Domestic Pressure: @Pluvio9yte contrasted OpenAI’s significant price cuts with the opposite move from Zhipu GLM, which increased its package prices several-fold. At the same time, @ruanyf analyzed that while Kimi K3’s performance is close to Fable 5, its high API pricing makes it one of the most expensive domestic models. @vista8, however, championed DeepSeek-V4-Flash, arguing its high cost-effectiveness represents “the AI that people can afford.”
• Interns vs. AI: The Replacement Competition: In insights shared by @Pluvio9yte on intern management, the fifth point bluntly states, “Most interns are not as good as Codex. If it weren’t for the fact that some tasks require a human, I would choose to buy more Codex licenses.” This sharply highlights the impact of AI programming tools on junior positions.

3. Recommended Tools and Resources Link to heading

📚 Appendix: Today’s Watch List Source Updates Link to heading

Timeframe: Last 3 days; 22 sources covered; 34 updates in total

Y Combinator Podcast (B_intro+search) Link to heading

• Patrick Collison: “What If You Succeed?”
  • Published: 2026-08-04 00:43 Beijing Time
  • Abstract: - You may have already heard of OpenClaw (formerly known as Clawdbot/Moltbot).
• The sensational open-source AI assistant that runs on your own device, connects with the messaging apps you already use, and goes beyond chat to actually perform tasks like managing email, calendars, files, workflows, and more.
• Now meet the person behind it.
• YC’s Raphael Schaad sat down with OpenClaw founder Peter Steinberger to talk about the “aha” moment behind the viral personal AI agent, why local-first agents could replace many of today’s apps, and how personal agents will reshape the future of software.
• EN Highlights:
  • In 2009, Patrick and John Collison went to Startup School in Berkeley, got sushi in Potrero Hill afterward, and decided on the walk home to start Stripe
  • The reasoning, as Patrick remembers it, was that “we might as well because it probably won’t be that hard.”
  • It took two years to launch
  • Seventeen years later, at Startup School 2026, he talks with YC’s Harj Taggar about dropping out of MIT twice, why founders should ask what happens if they succ…

Stratechery by Ben Thompson (A_full) Link to heading

• Meta Earnings, Meta’s Timing Problems, The Financial Tail
  • Published: 2026-08-03 18:00 Beijing Time
  • Summary: - Meta’s earnings were a bit disappointing; future promises about AI products were more disconcerting.
    • $15/month* or *$150/year.
    • Substantive analysis of the day’s news, delivered via three weekly emails or a podcast.
    • Strategy Interviews.
    • Interviews with leading public company CEOs, private company founders, and discussions with fellow analysts.
  • EN Highlights:
    • Meta’s earnings were a bit disappointing; future promises about AI products were more disconcerting.

OpenAI Blog (A_full) Link to heading

• How we built a realtime system for responsive voice AI in six months
  • Published: 2026-08-03 15:00 Beijing Time
  • Summary: - For voice AI, knowing when to speak is harder than it sounds.
    • Human speakers effortlessly switch turns in well under a second, but previous voice AI systems couldn’t keep up with this rhythm.
    • Their turn-based architectures relied on tiny models called turn detectors, which faced a tough task: guess too early, and the user gets cut off; guess too late, and the response is sluggish.
    • Only after the detector made its decision could the larger LLM get to work.
    • Its speech model is full-duplex, meaning it can listen and speak at the same time.
  • EN Highlights:
    • GPT-Live enables continuous voice interaction with AI, using a turnless speech model and low-latency architecture for faster, more natural conversations.

Two Minute Papers (B_intro+search) Link to heading

• Another DeepSeek Moment Has Arrived
  • Published: 2026-08-03 17:47 Beijing Time
  • Summary: - ❤️ Check out Lambda and sign up for their GPU Cloud here:.
• Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi.
• Another DeepSeek moment has arrived.
• EN Key Points:
  • ❤️ Check out Lambda here and sign up for their GPU Cloud:
  • 📝 DeepSeek v4 Flash 0731:
  • DeepSeek API:
  • 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:

ArXiv cs.AI (B_intro+search) Link to heading

• OpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28629v1 Announcement Type: new.
    • Abstract: The rapid transition from reactive large language models (LLMs) to persistent, action-capable systems has exposed critical gaps in the architectural understanding of Agentic AI, particularly in separating the reasoning, orchestration, and execution layers of autonomous AI agents.
    • Despite recent advances, unified frameworks for designing and evaluating full-stack agentic systems remain limited.
    • This paper proposes a comprehensive, layered Agentic AI architecture, outlining the evolution from reactive LLM interfaces to persistent, goal-driven autonomous AI agents with memory, planning, and continuous execution capabilities.
  • EN Key Points:
    • arXiv:2607.28629v1 Announce Type: new
    • Abstract: The rapid transition from reactive large language models (LLMs) to persistent, action-capable systems has exposed critical gaps in the architectural u…
    • Despite recent advances, unified frameworks for designing and evaluating full-stack agentic systems remain limited
    • This paper presents a comprehensive, layered architecture for Agentic AI, outlining the evolution from reactive LLM interfaces to persistent, goal-driven autono…

• Can AI Evaluate AI Scientists? A Benchmarking Study of Autonomous Research Generation Systems Using Automated Multi-Model Review

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28631v1 Announcement Type: new.
    • Abstract: AI scientist systems capable of autonomous research have the potential to significantly accelerate scientific discovery.
    • However, evaluating and comparing the quality of AI-generated papers remains an open challenge.
    • We propose and implement a rigorous benchmarking protocol using an automated peer review system that leverages cutting-edge large language models to evaluate scientific papers across four core dimensions: originality, scientific rigor, clarity, and significance.
  • EN Key Points:
    • arXiv:2607.28631v1 Announce Type: new

• Abstract: AI Scientist systems capable of autonomous research have the potential to significantly accelerate scientific discovery

  • However, evaluating and comparing the quality of AI-generated papers remains an open challenge
  • We propose and implement a rigorous benchmarking protocol using an automated peer-review system that harnesses frontier large language models to assess scientif…

• LLM Framework for Discovering Major Mathematical Conjectures: AI’s Quest for the Next Riemann Hypothesis

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28632v1 Announcement Type: New.
    • Abstract: Major mathematical conjectures still depend heavily on expert intuition, so there is still no unified method for systematically generating and validating conjectures with great mathematical potential.
    • We propose a three-stage pipeline for major conjecture discovery, including regional searches from explicit local evidence modules, reflective validation for foundational properties, novelty, and potential significance, and formal verification in Lean 4 and Mathlib.
    • The goal is to discover mathematical problems with high “problem taste”—that is, problems whose proofs could reorganize the language of a research field and provide lasting assistance to human mathematical research.
  • EN Key Points:
    • arXiv:2607.28632v1 Announce Type: new
    • Abstract: Major mathematical conjectures still depend heavily on expert intuition, so a unified method for the systematic generation and validation of conjectur…
    • We present a three stage pipeline for major conjecture discovery, with region search from explicit local evidence modules, reflective validation for foundationa…
    • The objective is the discovery of mathematical problems with high problem taste, namely problems whose proofs could reorganize the language of a research area a…

• ThinkReset: Learnable Intermediate Interface Construction for Bounded-Context Long-Horizon Reasoning

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28642v1 Announcement Type: New.
    • Abstract: Long-chain-of-thought reasoning improves performance on complex problems but also introduces redundancy accumulation, context overflow, and error anchoring.
    • We argue that under a bounded context window, the core bottleneck is not trajectory compression or test-time control, but the lack of a reusable intermediate interface to replace discarded history and support continued problem-solving.
    • We further identify a key failure mode for outcome-reward-driven long-chain reinforcement learning: when the model has not yet solved the task before the window is nearly exhausted, the final answer reward encourages premature guessing instead of continuing with careful reasoning.
  • EN Key Points:
    • arXiv:2607.28642v1 Announce Type: new

• Long-Chain-of-Thought for Inductive Reasoning

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28655v1 Announce Type: new.
    • Abstract: Long chain-of-thought reasoning improves performance on complex problems, but it also introduces redundancy accumulation, context overflow, and error…
    • We argue that under bounded context windows, the core bottleneck is not trajectory compression or test-time control, but the absence of a reusable intermediate…
    • We further identify a key failure mode of outcome-reward-driven long-chain reinforcement learning: when the model has not solved the task before the window is n…
  • EN Highlights:
    • Abstract: Long chain-of-thought reasoning improves performance on complex problems, but it also introduces redundancy accumulation, context overflow, and error…
    • We argue that under bounded context windows, the core bottleneck is not trajectory compression or test-time control, but the absence of a reusable intermediate…
    • We further identify a key failure mode of outcome-reward-driven long-chain reinforcement learning: when the model has not solved the task before the window is n…

• TAPR: Enhancing LLM Performance with a Task-Aware Prompt Rewriter

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28657v1 Announce Type: new.
    • Abstract: Large Language Models (LLMs) often require carefully designed prompts to unlock their full potential, which can be a barrier for non-expert users.
    • This work addresses this challenge by introducing the Task-Aware Prompt Rewriter (TAPR), a model that reformulates user prompts into task-optimized prompts with the explicit goal of improving downstream LLM performance.
    • We train TAPR using reinforcement learning with Group Relative Policy Optimization (GRPO), where rewards are derived from LLM-as-judge evaluations of both the reformulated prompts and the corresponding task outputs.
  • EN Highlights:
    • arXiv:2607.28657v1 Announce Type: new
    • Abstract: Large Language Models (LLMs) often require carefully crafted prompts to unlock their full potential, which can be a barrier for non-expert users
    • This work addresses the challenge by introducing a Task-Aware Prompt Rewriter (TAPR), a model that reformulates user prompts into task-optimized prompts with th…
    • We train TAPR using reinforcement learning with Group Relative Policy Optimization (GRPO), where rewards are derived from LLM-as-judge evaluations of both the r…

• Empowering Cross-Domain Sequential Recommendation with Hybrid Tokenization and Serial-Parallel Decoding

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28659v1 Announce Type: new.
    • Abstract: Cross-domain sequential recommendation (CDSR) aims to model users’ dynamic interest transitions and sequential patterns across multiple domains.
    • Recently, generative recommendation (GR) has emerged.
    • It first learns semantic identifiers (SIDs) from item semantics and formulates the recommendation as an autoregressive generation.
  • EN Highlights:
    • arXiv:2607.28659v1 Announce Type: new
    • Abstract: Cross-domain sequential recommendation (CDSR) aims to model users’ dynamic interest transitions and sequential patterns across multiple domains

• Recently, generative recommendation (GR) has emerged

• It first learns semantic identifiers (SIDs) from item semantics and formulates recommendation as autoregressive generation

• An Ontology-Guided, Deduplication-Aware Extraction Layer for Knowledge Graph Construction from Heterogeneous Documents

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28662v1 Announcement Type: New.
    • Abstract: Large language models extract entities and relationships from unstructured documents fluently but inconsistently: type vocabularies fracture across documents, the same person appears under multiple name variants, relationships are duplicated, and different individuals with shared names risk silent merging.
    • This paper presents the design, implementation, and empirical refinement of a production extraction layer that converts a live document stream into a validated knowledge graph aligned with a formal ontology.
    • The system consumes document metadata from Kafka, routes PDF, spreadsheet, Office, and image content through handlers built for each format, and extracts entities and relationships in two passes using a locally hosted Qwen3.5-9B model tuned on the ontology.
  • EN Key Points:
    • arXiv:2607.28662v1 Announce Type: new
    • Abstract: Large language models extract entities and relationships from unstructured documents fluently but inconsistently: type vocabularies fracture across do…
    • This paper presents the design, implementation, and empirical refinement of a production extraction layer that converts a live document stream into a validated…
    • The system consumes document metadata from Kafka, routes PDF, spreadsheet, Office, and image content through handlers built for each format, and extracts entiti…

• How Hard Does It Think? Analyzing Step-Aware Reasoning Energy in LLM Chain-of-Thought Trajectories

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28674v1 Announcement Type: New.
    • Abstract: Understanding how computational effort is allocated across individual Chain-of-Thought (CoT) reasoning steps remains an open challenge: existing interpretability methods rely on output-level signals or collapse processing depth into single trajectory-level scalars, rendering step-by-step workload opaque.
    • We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies workload at individual CoT step granularity via Centered Kernel Alignment (CKA) between Gram matrices of token hidden states across adjacent transformer layers, capturing inter-token relational structure without requiring feature vector alignment or cluster correspondence.
    • SARE further grounds this energy in the semantic progression of reasoning by modeling CoT trajectories as transitions between latent semantic states.
  • EN Key Points:
    • arXiv:2607.28674v1 Announce Type: new
    • Abstract: Understanding how computational effort is allocated across individual chain-of-thought (CoT) reasoning steps remains an open challenge: existing inter…

• We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies effort at the granularity of individual CoT steps via Centered Kernel Align…

• SARE further contextualizes this energy within reasoning’s semantic progression by modeling CoT trajectories as transitions among latent semantic states

• Reasoning in Real World Clinical Care: Why Large Language Models Are Not Yet Safe for Autonomous Clinical Decision Support

  • Published: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28677v1 Announcement Type: New.
    • Abstract: LLMs can now pass medical licensing exams and, in carefully curated cases, can rival physicians in diagnostic reasoning.
    • These developments have accelerated the use of LLMs for symptom assessment and clinical decision support in diagnostic and treatment guidance, administrative documentation, and rule-based alert enhancement.
    • This perspective addresses the most significant of these applications: the autonomous triage of self-presenting, undifferentiated patients with little to no clinician involvement.
  • EN Highlights:
    • arXiv:2607.28677v1 Announce Type: new
    • Abstract: LLM now pass medical licensing examinations and, in curated cases, can rival physicians at diagnostic reasoning
    • These developments have accelerated the use of LLMs for symptom assessment and clinical decision support in diagnostic and treatment guidance, administrative do…
    • This Perspective concerns the most consequential of these applications: the autonomous triage of self-presenting, undifferentiated patients, with little or no c…

• ViSAGE: Constructing Self-Correcting Memories for Long-Form Video Understanding

  • Published: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28678v1 Announcement Type: New.
    • Abstract: Multimodal agents operating in long-horizon environments must construct and continuously update multimedia memories to support entity-consistent, time-based reasoning.
    • However, existing agentic memory methods often discard fine-grained identity cues under aggressive compression and segmented processing.
    • They also heavily rely on vector similarity retrieval, which can present semantically relevant but identity-mismatched evidence, leading to entity confusion, error propagation, and hallucinatory answers.
  • EN Highlights:
    • arXiv:2607.28678v1 Announce Type: new
    • Abstract: Multimodal agents operating in long-horizon environments must build and continually update multimedia memories to support entity-consistent, temporall…
    • However, existing agentic memory approaches often discard fine-grained dentity cues under aggressive compression and segment-wise processing

• They also rely heavily on vector similarity retrieval, which can surface semantically related yet identity-mismatched evidence, leading to entity confusion, err…

ArXiv cs.CL (B_intro+search) Link to heading

• Can LLMs Really Understand Item Difficulty Levels? Implications for Automated Item Generation Using LLMs

  • Published: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28634v1 Announce Type: new.
    • Abstract: The estimation of item difficulty plays a key role in both formative assessment and large-scale high-stakes summative assessments.
    • This study explores how large language models (LLMs) perform in predicting item difficulty levels using items from a large-scale Reading and Writing test.
    • The study investigated various prompting strategies and parameter settings across multiple LLMs.
  • EN Highlights:
    • arXiv:2607.28634v1 Announce Type: new
    • Abstract: The estimation of item difficulty plays a key role in both formative assessment and large-scale high-stakes summative assessments
    • This study explores how large language models (LLMs) perform in predicting item difficulty levels using items from a large-scale Reading and Writing test
    • The study investigated various prompting strategies and parameter settings across multiple LLMs

• Imbalanced Data Clustering via Targeted Data Augmentation Using GMM and LLM

  • Published: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28635v1 Announce Type: new.
    • Abstract: In Natural Language Processing (NLP), dealing with underrepresented topics is challenging, especially in unsupervised tasks where clustering may not adequately capture minority topics.
    • To address this challenge, our paper proposes a novel unsupervised data augmentation method that integrates Gaussian Mixture Models (GMMs) and Large Language Models (LLMs).
    • Due to their flexibility and robustness, GMMs can detect clusters corresponding to underrepresented areas in the data, while LLMs create synthetic documents to enrich these clusters and improve their representation.
  • EN Highlights:
    • arXiv:2607.28635v1 Announce Type: new
    • Abstract: In Natural Language Processing (NLP), dealing with underrepresented topics is challenging, especially in unsupervised tasks where clustering might not…
    • To tackle this challenge, our paper presents a novel unsupervised data augmentation method that integrates Gaussian Mixture Models (GMMs) and Large Language Mod…
    • Due to their flexibility and robustness, GMMs can detect clusters corresponding to underrepresented areas in the data, while LLMs create synthetic documents to…

• Chain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28636v1 Announcement Type: New.
    • Abstract: LLMs increasingly serve as automated judges, but their judgments remain vulnerable to cognitive biases.
    • Existing mitigations mostly rely on prompt-driven debiasing, which is brittle across bias types, or human evaluation, which does not scale.
    • We study \emph{Chain-of-Models} (CoM), an automated audit pipeline in which a second model inspects the first model’s reasoning trace before producing the final judgment.
  • EN Key Points:
    • arXiv:2607.28636v1 Announce Type: new
    • Abstract: LLMs increasingly serve as automated judges, but their judgments remain vulnerable to cognitive biases
    • Existing mitigations mostly rely on prompt-driven debiasing, which is brittle across bias types, or human evaluation, which does not scale
    • We study \emph{Chain-of-Models} (CoM), an automated audit pipeline in which a second model inspects the first model’s reasoning trace before producing the final…

• ZeroR @CHiPSAL 2026: Two-Stage Vision-Language Adaptation with Contrastive Learning for Nepali Meme Classification

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28637v1 Announcement Type: New.
    • Abstract: This paper presents our CHiPSAL 2026 shared task system, which involves multimodal hate speech and sentiment detection in Nepali memes.
    • We address two subtasks: binary hate speech classification and three-class sentiment analysis.
    • Our method uses Qwen3-VL-8B-Instruct to adapt the Robust Adaptation of Hateful Meme Detection (RA-HMD) framework, Qwen3-VL-8B-Instruct is a state-of-the-art vision-language model with native Sanskrit support.
  • EN Key Points:
    • arXiv:2607.28637v1 Announce Type: new
    • Abstract: This paper presents our system for the CHiPSAL 2026 shared task on multimodal hate speech and sentiment detection in Nepali memes
    • We address both subtasks: binary hate speech classification and three-class sentiment analysis
    • Our approach adapts the Robust Adaptation of Hateful Meme Detection (RA-HMD) framework using Qwen3-VL-8B-Instruct, a state-of-the-art vision-language model with…

• Learning Stateful Predictive Knowledge From Experience

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28638v1 Announcement Type: New.
    • Abstract: As Large Language Model (LLM) agents increasingly learn from experience, they primarily rely on trajectory-level reflection to extract insights.
    • From the perspective of predictive knowledge, we argue that this approach is based on episodic hindsight rather than predictive foresight, leading to fragile, path-dependent heuristics.
    • To address this issue, we propose Stateful Knowledge Learning (SKL).

• EN Key Points:

  • arXiv:2607.28638v1 Announce Type: new
  • Abstract: As large language model (LLM) agents increasingly learn from experience, they primarily rely on trajectory-level reflection to extract insights
  • Viewed through the lens of predictive knowledge, we argue that this approach operates on episodic hindsight rather than predictive foresight, yielding brittle,…
  • To address this, we propose Stateful Knowledge Learning (SKL)

• The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28639v1 Announce Type: new.
    • Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias.
    • On unambiguous tasks (BBQ-disambig), response-based distillation from a Gemma-2-9B teacher improves context-following: for the most biased baseline (SmolLM2-1.7B-Instruct), it reduces the context coverage error rate from 44% to 24%.
    • On ambiguous tasks (BBQ-ambig), the same distillation destroys per-item refusal calibration: 15% of items where the baseline correctly abstained instead received a stereotypical answer.
  • EN Key Points:
    • arXiv:2607.28639v1 Announce Type: new
    • Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias
    • On unambiguous tasks (BBQ-disambig), response-based distillation from a Gemma-2-9B teacher improves context-following: for the most biased baseline (SmolLM2-1.7…
    • On ambiguous tasks (BBQ-ambig), the same distillation destroys per-item refusal calibration: 15% of items where the baseline correctly abstained instead receive…

• TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28640v1 Announce Type: new.
    • Abstract: Multimodal large language models (MLLMs) should generate consistent responses given semantically equivalent inputs across modalities.
    • However, we observe a systematic discrepancy in model predictions under such cross-modal variations.
    • Specifically, we define the modality gap as the difference in model performance under semantically equivalent text and multimodal inputs.
  • EN Key Points:
    • arXiv:2607.28640v1 Announce Type: new
    • Abstract: Multimodal large language models (MLLMs) should generate consistent responses given semantically equivalent inputs across modalities
    • However, we observe a systematic discrepancy in model predictions under such cross-modal variations

• Specifically, we define the modality gap as the difference in model performance under semantically equivalent textual and multimodal inputs

• The Formalism Trap: Are LLM-as-a-Judge Evaluators Blinded by Consensus Mimicry under Social Load?

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28641v1 Announcement Type: New.
    • Abstract: We introduce the \textit{Agentic Formalism Trap} and the Evaluative Dissonance Index ($D_E$), quantifying how LLM-as-a-Judge systems conflate structural proceduralism with semantic truth under adversarial loads.
    • By analyzing 22,500 trajectories across 3 domains (GAIA, SWE-bench, Multi-Challenge), we extract a semantic taxonomy of hallucination maneuvers, validated through a deterministic lexical basis ($p < 10^{-120}$).
    • A logistic meta-evaluator isolates the exact syntactic triggers captured by this evaluator (ROC-AUC 0.8779), while zero-shot Leave-One-Domain-Out transfer proves this vulnerability is universally domain-agnostic (average ROC-AUC 0.7482).
  • Key Points:
    • arXiv:2607.28641v1 Announce Type: new
    • Abstract: We introduce the \textit{Agentic Formalism Trap} and the Evaluative Dissonance Index ($D_E$), quantifying how LLM-as-a-Judge systems conflate structur…
    • Analyzing 22,500 trajectories across 3 domains (GAIA, SWE-bench, Multi-Challenge), we extract a semantic taxonomy of hallucination maneuvers, validated via dete…
    • A logistic meta-evaluator isolates the exact syntactic triggers of this evaluator capture (ROC-AUC 0.8779), while a zero-shot Leave-One-Domain-Out transfer prov…

• Evaluating Federated Pre-Training: On the Reliability of Downstream Fine-Tuning and Intrinsic Evaluation

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28658v1 Announcement Type: New.
    • Abstract: Federated pre-training offers a method to train foundation models on private or distributed data without centralizing the underlying datasets.
    • However, evaluating federated pre-training remains challenging because differences in client participation and local data availability can make directly comparable evaluations difficult.
    • Furthermore, pre-training test perplexity is related to the pre-training distribution, and downstream benchmarks introduce task-specific adaptations that may not faithfully reflect the test perplexity established during pre-training.
  • Key Points:
    • arXiv:2607.28658v1 Announce Type: new
    • Abstract: Federated pre-training offers a way to train foundation models on private or distributed data without centralizing the underlying datasets
    • However, evaluating federated pre-training remains challenging because differences in client participation and local data availability can make directly compara…

• Moreover, pre-training test perplexity is tied to the pre-training distribution, while downstream benchmarks introduce task-specific adaptation that may not fai…

• Are the Financial Reasoning from LLMs Credible? A Real World Test over Long-Horizon Statements

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28661v1 Announcement Type: New.
    • Abstract: Do Large Language Models (LLMs) possess genuine structural reasoning, or do they merely rely on surface-level pattern matching?
    • The financial domain, which requires numerical precision and multi-step logic in long-term contexts, is an ideal testbed.
    • Existing benchmarks fail to capture real-world industrial complexity, primarily relying on multiple-choice questions or single-hop QA on cropped tables, while ignoring complex cross-statement dynamics and temporal accumulation.
  • EN Key Points:
    • arXiv:2607.28661v1 Announce Type: new
    • Abstract: Do Large Language Models (LLMs) possess genuine structural reasoning, or merely rely on surface-level pattern matching
    • The financial domain, demanding numerical precision and multi-step logic over long contexts, is an ideal testbed
    • Existing benchmarks fail to capture real-world industrial complexity, predominantly relying on multiple-choice questions or single-hop QA over cropped tables wh…

ArXiv cs.LG (B_intro+search) Link to heading

• Topology-Aware Data Movement for Disaggregated GPU Inference

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28633v1 Announcement Type: New.
    • Abstract: Disaggregated LLM inference creates a datacenter networking problem that existing systems cannot solve correctly.
    • When pre-filling and decoding run on separate GPU pools, the KV cache must be transferred between them.
    • For a 70B model, this amounts to 2.6 GB per request, totaling over 100 GB/s at production scale.
  • EN Key Points:
    • arXiv:2607.28633v1 Announce Type: new
    • Abstract: Disaggregated LLM inference creates a datacenter networking problem that no existing system solves correctly
    • When prefill and decode run on separate GPU pools, the KV cache must be transferred between them
    • For a 70B model this is 2.6 GB per request, exceeding 100 GB/s aggregate at production scale

• Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28665v1 Announcement Type: New.
    • Abstract: Automated Driving Systems (ADS) are becoming ubiquitous.

• Future Software Defined Vehicles (SDVs) may be able to run multiple ADSs, both native and aftermarket, such as Comma.ai’s Openpilot.

• Monitoring systems to independently verify which automated driving system is active are important for safety monitoring, regulatory compliance, insurance assessment, and anomaly detection.

  • EN Highlights:
    • arXiv:2607.28665v1 Announce Type: new
    • Abstract: Automated driving systems (ADSs) are becoming ubiquitous
    • Future Software Defined Vehicles (SDVs) may be able to run multiple ADSs, both native and aftermarket such as Comma.ai’s Openpilot
    • Monitoring systems to independently verify which automated driving system is active are important for safety monitoring, regulatory compliance, insurance assess…

• Guarantees on Dynamical System Distinguishability for LLM Token Generation

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28667v1 Announce Type: new.
    • Abstract: Recent work has shown that large language model (LLM) responses can be distinguished by modeling token embeddings as trajectories of a black-box dynamical system (DS) and comparing the prediction residuals of the two DSs.
    • Despite the empirical success of this dynamical approach, a theoretical understanding is still lacking as to why it works, how well it scales as a function of the token sequence, and when it transfers across embedding models.
    • We address these questions by formalizing the classification task as a binary hypothesis test between two stochastic linear DSs.
  • EN Highlights:
    • arXiv:2607.28667v1 Announce Type: new
    • Abstract: Recent work has shown that classifying large language models (LLMs)’ responses can be distinguished by modeling token embeddings as trajectories of a…
    • Despite the empirical success of this dynamical approach, a theoretical understanding of why it works, how well it scales as a function of the token sequence, a…
    • We address these questions by formalizing the classification task as a binary hypothesis test between two stochastic linear DSs

• LARA: Lightweight Adapters in the Residual Stream for Composable Adaptation and Alignment

  • Release Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28669v1 Announce Type: new.
    • Abstract: We propose LARA (Lightweight Additive Residual Adaptation), an efficient adaptation method that operates in the residual stream of a frozen model rather than in its weights.
    • Whereas LoRA adds low-rank updates to weight matrices, LARA reads the hidden states of a small set of layers and adds a low-rank correction back into the residual stream, leaving all base weights untouched.
    • On code fine-tuning tasks and with preference optimization (DPO), LARA matches LoRA at the same parameter counts.
  • EN Highlights:
    • arXiv:2607.28669v1 Announce Type: new

• Abstract: We present LARA (Lightweight Additive Residual Adaptation), a method for efficient adaptation that operates in the residual stream of a frozen model r…

• Where LoRA adds an update of low rank to weight matrices, LARA reads the hidden state at a small set of layers and adds a correction of low rank back to the res…

• On a code fine-tuning task and on preference optimization (DPO), LARA matches LoRA at equal parameter counts

• Hierarchical Copula-Gumbel-Top-\texorpdfstring{$K$}{K} Routing: Two-Sided Dependence Control for Frozen Mixture-of-Experts at Fixed Per-Token Routing Laws

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28670v1 Announce Type: new.
    • Abstract: A stochastic Gumbel-Top-$K$ router defines, for every token of a Mixture-of-Experts (MoE) model, a \emph{routing law}: a distribution over ordered lists of experts and mixing weights.
    • We ask which \emph{joint} distributions over the routing choices of different tokens are reachable while every individual token’s complete routing law is held entirely fixed.
    • We give a two-sided construction, \emph{Hierarchical Copula-Gumbel-Top-$K$} (\CGA{}).
  • EN Highlights:
    • arXiv:2607.28670v1 Announce Type: new
    • Abstract: A stochastic Gumbel-Top-$K$ router defines, for every token of a mixture-of-experts (MoE) model, a \emph{routing law}: a distribution over ordered exp…
    • We ask which \emph{joint} distributions over the routing choices of different tokens are reachable while every individual token’s complete routing law is held e…
    • We give a two-sided construction, \emph{Hierarchical Copula-Gumbel-Top-$K$} (\CGA{})

• LAWFUL: Law-Aligned Witness for Faithful Use of Latents

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28672v1 Announce Type: new.
    • Abstract: When a neural network accurately predicts a physical system, does it learn the governing law as formal, structured knowledge? If so, do the network’s internal computations actually use that representation across the law’s full domain of validity?
    • We identify four explainability gaps that have limited answering these questions for {\em physical laws over continuous variables}: a lack of coverage-aware causal consistency metrics for continuous counterfactuals; validity domain testing for identified circuits; verifying a law’s invariances and forbidden behaviors; and quantifying how derived physical quantities flow through a circuit.
    • We develop a foundational framework, LAWFUL, that closes the first two and builds foundations for the remaining two, and illustrate it on the Mocap2Radar transformer, verifying that it learns and internally uses the Doppler frequency law $f(t) = \frac{2 v(t)}{\lambda}$ from motion capture and radar data, where neither $f(t)$ nor $v(t)$ appears.
  • EN Highlights:
    • arXiv:2607.28672v1 Announce Type: new

• Abstract: When a neural network predicts a physical system accurately, has it learned the governing law as formal, structured knowledge, and if so, does the net…

• We identify four interpretability gaps that limit answering these questions for {\em physics laws over continuous variables}: the absence of a coverage-aware ca…

• We develop a foundational framework, LAWFUL, that closes the first two and lays groundwork for the remaining two, and illustrate it on the Mocap2Radar transform…

• MPP-GNN: Subject-Adaptive Community Detection for fMRI-Based Alzheimer’s Disease Classification

  • Published: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28681v1 Announce Type: new.
    • Abstract: Functional magnetic resonance imaging (fMRI) is a widely used technique for studying the brain.
    • Recent methods utilizing graph neural networks (GNNs) to analyze brain functional connectivity have shown great potential in the classification of brain diseases such as Alzheimer’s disease (AD).
    • However, these methods often assume a preset number of functional modules for all subjects, which overlooks inter-subject variability.
  • EN Key Points:
    • arXiv:2607.28681v1 Announce Type: new
    • Abstract: Functional magnetic resonance imaging (fMRI) is a widely used technique for studying the brain
    • Recent methods that utilize graph neural networks (GNNs) for analysis of brain functional connectivity have shown great potential for the classification of brai…
    • However, these methods often assume a preset number of functional modules across all subjects, which overlooks inter-subject variability

• Technological Advances in Detecting and Managing Cognitive Impairment in Older Adults: Trends, Challenges, and Future Directions

  • Published: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28687v1 Announce Type: new.
    • Abstract: As the population ages, cognitive decline from mild cognitive impairment (MCI) to dementia is a defining health challenge for the coming decades, yet routine assessments often miss its earliest signs.
    • This paper critically synthesizes the latest technological advances in detecting and managing cognitive impairment in older adults, covering neurophysiological signals (primarily electroencephalography, EEG), structural and molecular neuroimaging (MRI and amyloid/tau PET), blood biomarkers, and digital markers, integrated via artificial intelligence (AI), machine learning (ML), and deep learning (DL).
    • In addition to a summary, it provides an interdisciplinary taxonomy, a lens for methodological rigor, subject- and location-independent validation, a comprehensive early detection framework that links tiered screening with interventions, and comparative tables of detection methods, interventions, and risk and protective factors.
  • EN Key Points:
    • arXiv:2607.28687v1 Announce Type: new

• Abstract: As populations age, cognitive decline from mild cognitive impairment (MCI) to dementia is a defining health challenge of the coming decades, yet routi…

  • This article critically synthesizes recent technological advances for detecting and managing cognitive impairment in older adults, spanning neurophysiological s…
  • Beyond summarizing, it contributes a cross-disciplinary taxonomy, a methodological-rigor lens foregrounding subject- and site-independent validation, an integra…

• SEDR-Seq2P: A Lightweight Dilated Residual Sequence-to-Point Network for Multi-Task Industrial NILM

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28693v1 Announcement Type: new.
    • Abstract: Industrial NILM remains challenging because measurement noise and widespread concurrent machine operation reduce the generalization of models tuned on residential data.
    • This work adopts a one-to-many, multi-task disaggregation setting, in which a single network estimates multiple industrial machine loads from aggregate power.
    • Under a unified evaluation protocol on IMDELD, we benchmark Seq2Seq, Seq2SubSeq, Seq2Point, GRU, and WaveNet using energy-estimation metrics and the accuracy-delay criterion.
  • EN Highlights:
    • arXiv:2607.28693v1 Announce Type: new
    • Abstract: Industrial NILM remains challenging because measurement noise and widespread concurrent machine operation reduce the generalization of models tuned on…
    • This work adopts a one-to-many, multi-task disaggregation setting, in which a single network estimates multiple industrial machine loads from aggregate power
    • Under a unified evaluation protocol on IMDELD, we benchmark Seq2Seq, Seq2SubSeq, Seq2Point, GRU, and WaveNet using energy-estimation metrics and the accuracy-de…

• Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning

  • Publication Time: 2026-08-03 12:00 Beijing Time
  • Abstract: - arXiv:2607.28695v1 Announcement Type: new.
    • Abstract: This is a plain text version optimized for the arXiv submission form.
    • Custom macros (such as \CV and \SI) have been converted to standard text/math so that they render correctly on the webpage: Evaluating the fatigue life of structural steels typically requires mechanical tests lasting tens to hundreds of hours, making rapid quality control impractical.

• We propose CV, a computer vision framework that directly estimates the fatigue life ($\log N_f$) of lightweight alloy steels from optical micrographs, eliminating the need for physical testing. The process features a seven-stage OpenCV preprocessing routine for artifact removal, a 28-dimensional physics-informed feature extractor (quantifying crack morphology, grain structure, porosity, and texture), and a CNN regression model trained with Gaussian Negative Log-Likelihood (GNLL) loss to jointly predict $\log N_f$ and sample-specific uncertainty $\hat{\sigma}$. Evaluating three architectures (SE-CNN, ResNet-50, VGG-16) on a synthetic micrograph benchmark, ResNet-50 achieved an $R^2 = 0.93$, RMSE = 0.18 log cycles, and a macro F1 = 0.91.

  • EN Highlights:
    • arXiv:2607.28695v1 Announce Type: new
    • Abstract: Here is the plain text version optimized for arXiv’s submission form
    • Custom macros (like \CV and \SI) have been converted to standard text/math so they render correctly on the webpage: Evaluating the fatigue life of structural st…
    • We present CV, a computer vision framework that estimates the fatigue life ($\log N_f$) of lightweight alloy steels directly from optical micrographs without ph…