2026-06-26 AI Daily | Growing Focus on Model Controllability, Agents Move from Demos to Verifiable Delivery Link to heading

Today’s focus shifts from model capabilities to behavioral boundaries and engineering delivery. Multiple studies indicate that alignment is jointly influenced by data, persona settings, and post-training processes. Agent evaluation is beginning to focus on efficiency, reliability, and chains of responsibility. On the enterprise side, persistent collaborative agents, open-source code models, and AI observability are advancing in parallel, but risks related to privacy, permissions, and autonomous execution still require careful handling.

📖 In-Depth Guide to this Issue’s Watch List Link to heading

The most noteworthy topic today is a set of papers on “model behavioral controllability”: from the cascading linear features of sycophancy, to how persona affects refusal, and how post-training processes might weaken compassionate values, several studies converge on one issue—alignment is not a single switch but is shaped collectively by data, role-setting, and downstream behavior, making them a must-read for security and evaluation teams.

The second main theme is “agent evaluation and governance.” The CORE-Bench paper reminds us that after benchmarks saturate, we should not just chase harder problems but also look at dimensions like efficiency, reliability, and human-computer collaboration. The validation dilemma of Coding Agent rewards, and the institutional proof model of “governing actions, not agents,” are also shifting the focus from capability demonstrations to verifiable and accountable deployments.

Finally, on the application side, it’s worth noting the exploration of knowledge-enhanced agents in mental health medication information, as well as directions like LLM-driven algorithmic trading and DAO/enterprise AI protocol governance analysis. These show that agents are entering high-risk, high-complexity scenarios, but the real barriers remain evidence boundaries, accountability structures, and continuous validation.

🌐 Breaking AI News on X Link to heading

Topic 1: Anthropic Launches Claude Tag for Persistent Slack AI Teammate Link to heading

• Category: AI · News
• Overview: Trending for 2 days, 53,000 related posts
• What it is: Anthropic has launched Claude Tag, which positions Claude as an AI teammate in Slack that is persistently online, remembers context, and assists with workflows. It is aimed at Enterprise/Team plans and replaces the old Slack app.
• Why it matters: This marks a shift for enterprise AI from passive Q&A tools to collaborative agents with long-term memory, environmental awareness, and autonomous execution capabilities, potentially changing how teams manage knowledge, develop software, and collaborate daily.
• Discussion overview: Discussions on X focus on whether Claude Tag can truly reduce context loss, productize tacit enterprise knowledge, and which workflows are suitable for a persistent AI. The debate centers on the balance between the efficiency gains from its autonomous monitoring of Slack messages and the associated risks to privacy, permissions, and reliability.

Topic 2: GLM-5.2 Emerges as Top Open-Weight Coding Model Link to heading

• Category: AI · News
• Overview: Trending for 22 hours, 529 related posts
• What it is: Zhipu AI’s GLM-5.2 is being widely discussed by X users as one of the top-performing open-weight code models available.
• Why it matters: This indicates that open-weight models are continuing to close the gap with cutting-edge closed-source models in code generation and software engineering tasks. It could lower the cost for businesses and developers to use high-performance programming AI and accelerate localized, controllable deployments.
• Discussion overview: Discussions on X center on GLM-5.2’s actual benchmark performance, its cost-effectiveness compared to models from Google, MiniMax, etc., and whether open-weight models will further transform software development workflows in 2026. The main points of contention are whether its capabilities have reached a ‘cutting-edge’ level and if benchmark results can translate into stable, real-world development efficiency.

Topic 3: Sazabi Raises $8M to Build Self-Healing AI Observability Platform Link to heading

• Category: AI · News
• Overview: Trending for 23 hours, 1,400 related posts
• What it is: AI observability startup Sazabi has raised $8 million in funding to build an AI system monitoring and operations platform with ‘self-healing’ capabilities.
• Why it matters: As enterprises deploy more large model applications, it is becoming harder to promptly detect model anomalies, performance degradation, cost overruns, and security risks. AI observability and automated remediation capabilities are becoming a critical part of production-grade AI infrastructure.
• Discussion overview: Discussions on X mainly focus on whether ‘self-healing’ AI operations can genuinely reduce manual troubleshooting costs and whether this sector will become the next infrastructure hotspot after model development platforms. Some also question if the early-stage product capabilities are being exaggerated by the funding narrative.

Summary of Today’s AI Sentiment on X Link to heading

The main theme in today’s public discourse is the shift of AI from a “model capability race” to “real-world production deployment”: on one hand, persistent enterprise collaboration agents like Claude Tag are being integrated into Slack workflows; on the other, open-weight code models such as GLM-5.2 are lowering the barrier to entry for high-performance programming AI. Meanwhile, solutions like Sazabi, representing observability and self-healing operations, are starting to address the infrastructure shortcomings in post-deployment enterprise environments. The consensus is that AI is no longer just a tool for chat or single-point generation but is expanding towards long-term memory, contextual awareness, autonomous execution, locally controllable deployment, and automated operations. The main point of contention is whether a “perceived lead” can be consistently translated into real production efficiency: is Claude Tag worth the trade-off between collaboration efficiency and privacy/permission risks? Does GLM-5.2 truly achieve state-of-the-art coding capabilities? And have self-healing operations solutions like Sazabi reached sufficient product maturity? The potential risk is that enterprises are too quick to hand over critical knowledge flows, code processes, and operational judgments to AI, exposing new systemic issues related to permission boundaries, data leakage, erroneous autonomous execution, evaluation bubbles, and runaway costs.

💡 Influencer Insights Link to heading

As a senior AI industry analyst, I have compiled the following industry intelligence based on influencer perspectives on the X platform over the past 24 hours. Here is the core summary:

1. Key Technical Trends and Hot Products Watched by Influencers Today Link to heading

Core Trend: The Battle for “Agentic Operating Systems” and Production-Ready Engineering

• Product in Focus: The Ecological Niche and Bottlenecks of Codex

  • Repositioning Upwards: The industry is no longer satisfied with viewing Codex as merely a programming tool. @dotey clearly states that the development trend for Codex is to become an “Agent OS,” not just an “Agent Office.” He shared his experience decompiling and replicating Codex code projects, revealing that its underlying logic is being deeply studied by the community.
  • Usage and Cost Anxiety: Several bloggers (@Pluvio9yte, @ruanyf) reported that Codex’s token consumption has surged or its limits have become stricter, jokingly referring to it as a “shrinking multi-billion subsidy.” Simultaneously, @Pluvio9yte and @vista8 shared their hybrid workflows within Codex, which involve combining different models (e.g., using Gemini for chat and Claude for planning) to overcome the limitations of a single model.

• The Model Arena: A Tug-of-War Between On-Device, Coding, and Multimodal Capabilities

  • On-Device Models Become a Red Ocean: @zhixianio continues to conduct in-depth tests of on-device models, from Google’s Gemma series (4 12B Coder, E4B + MTP) to the audio-video full-duplex capabilities of MiniCPM-o 4.5, giving them high praise and considering them “ready for use.” The QAT (Quantization-Aware Training) model released by @googledevs is seen as a key optimization approach.
  • Real-World Code Generation Capability Tests: @zhixianio conducted a deep comparative review of Gemma 4 12B Coder and Qwen3.6-35B-A3B, pointing out that 12B models hit a ceiling when handling complex, stateful programs (like Tetris), while the 35B MoE remains the “sweet spot.” @Pluvio9yte shared a tutorial on integrating cost-effective models like Doubao Seed 2.1 Pro into Claude Code, signaling the diversification and impending price wars in the model supply market.
  • Lightweight AI Video Generation Tools: Topview AI’s Seedance 2.0 Mini has been launched, focusing on being “fast and cheap.” @AI_Jasonyu finds it very user-friendly for daily AI-powered comic creation, reflecting a shift in the AI video domain from competing solely on image quality to focusing on cost-effectiveness and speed.

2. Noteworthy Unique Perspectives or Industry Foresight Link to heading

• The “New Cold War” in AI: Distillation Attacks and Government Regulation

  • @dotey provided an in-depth analysis of the incident where Anthropic accused Alibaba of large-scale distillation of Claude, noting that the scale of the attack exceeded the total of all previous Chinese companies combined. He dissected Anthropic’s “awkward” position of both seeking government help and protesting government restrictions on model releases, pointing out that this reflects a direct collision between US-China AI capabilities and commercial interests.
  • He also reported on the unprecedented process for OpenAI’s GPT-5.6 release, which was changed to a “case-by-case customer approval” basis due to government requirements. He warned that this could widen the gap between the company’s internal capabilities and what is publicly available.

• A Shift in Thinking: From Vibe Coding to Agent Delivery

• After “Vibe Coding” Comes “Engineering”: @gefei55 pointed out that Tokens are infinite, but time and energy are limited, warning everyone not to get lost in the “Token trap of being able to do anything.” @Pluvio9yte also highlighted that in the AI era, anything repeated more than three times must be automated, and shared a case of using a “Skill Hygiene skill” to clean up outdated scripts.

• The “Last Mile” from Demo to Product: @AI_Jasonyu and @vista8 both promoted Agent deployment platforms like EdgeOne Makers, emphasizing that they solve pain points encountered when moving from a local setup to a live environment, such as concurrency, sandbox security, and memory storage. This signifies a shift in the industry’s focus from “Can it run?” to “Can it be delivered to real users?”

• On the Philosophy of AI Creation and Knowledge Management

  • The “Model Imprint” on Language Style: @lijigang made a profound observation: heavy use of a specific model can cause one to adopt its linguistic style, resulting in a “Claude flavor” or a “Deepseek flavor,” and noted that the brain’s neural networks are very “receptive” to Context.
  • “Machine-Readable” Formats for Knowledge Management: @vista8 introduced Google’s Open Knowledge Format (OKF), the core idea of which is to package knowledge using Markdown + YAML frontmatter into version-controllable file bundles that can be directly consumed by Agents. This suggests that optimizing knowledge structures for AI will become a new skill for individuals and organizations.

3. Recommended Tools and Resources Link to heading

📚 Appendix: Today’s Watch List Update Sources Link to heading

Timeframe: Last 3 days; 22 sources covered; 32 updates in total

Stratechery by Ben Thompson (A_full) Link to heading

• An Interview with Figma CEO Dylan Field About Design and AI
  • Publication Time: 2026-06-25 18:00 Beijing Time
  • Abstract: - Field is a Thiel Fellow who dropped out of Brown University in 2012 to found Figma.
• Figma has been on a fascinating journey: the company accepted an acquisition offer from Adobe in 2022, but regulatory resistance forced the latter to abandon the merger in late 2023.
• I talked with Field about all of this, including his background, Figma’s differentiated discovery process, and the nature of creativity and design.
• We discussed the issue of artificial intelligence, which the market sees as a headwind, but Field views as a tailwind.
• As a reminder, all Stratechery content (including interviews) is available as a podcast; click the link at the top of this email to add Stratechery to your podcast player.
• EN Highlights:
  • Listen to this post:
  • Good morning,
  • This week’s Stratechery interview is with Figma co-founder and CEO Dylan Field
  • Field was a Thiel Fellow who dropped out of Brown in 2012 to start Figma

OpenAI Blog (A_full) Link to heading

• How agents are transforming work
  • Posted: 2026-06-25 10:00 Beijing Time
  • Summary: - Agent AI transforms the unit of knowledge work from a single interaction to a delegated long-term task.
    • Chatbot interactions are typically short and self-contained.
    • Agents can operate independently for minutes or hours, while orchestrating tool calls, interacting with the environment, and iterating on solutions.
    • Consequently, agents are quickly becoming the most powerful AI tools for work.
    • Last year, we witnessed this transformation firsthand at OpenAI.
  • EN Highlights:
    • A new OpenAI research paper shows how AI agents are transforming work, enabling longer, more complex tasks and expanding productivity across roles.

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

• Detecting and Controlling Sycophancy with Cascading Linear Features

  • Posted: 2026-06-26 12:00 Beijing Time
  • Summary: - arXiv:2606.26155v1 Announce Type: new.
    • Abstract: Interpreting and controlling model behaviors through activation steering methods requires many pairs of contrastive samples that clearly exhibit desired or undesired behaviors.
    • These data pairs determine the degree to which interpretability frameworks can reliably detect model features responsible for a behavior, and therefore the ability to steer the model towards or away from that behavior.
    • In this work, we present an iterative data generation pipeline that isolates cascading linear features responsible for a behavior.
  • EN Highlights:
    • arXiv:2606.26155v1 Announce Type: new
    • Abstract: Interpreting and controlling model behaviors through activation steering methods requires many pairs of contrastive samples that clearly exhibit desir…
    • These data pairs determine the degree to which interpretability frameworks can reliably detect model features responsible for a behavior, and therefore the abil…
    • In this work, we present an iterative data generation pipeline that isolates cascading linear features responsible for a behavior

• Life After Benchmark Saturation: A Case Study of CORE-Bench

  • Posted: 2026-06-26 12:00 Beijing Time
  • Summary: - arXiv:2606.26158v1 Announce Type: new.

• Abstract: When a benchmark’s accuracy saturates, it is often retired and replaced with a more challenging version.

• We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues, such as shortcuts, out-of-distribution generalization, efficiency, reliability, the relative importance of the model versus the scaffold, and the gains from human-agent collaboration.

• We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimensions can yield meaningful insights into agent performance even after accuracy has saturated.

  • EN Highlights:
    • arXiv:2606.26158v1 Announce Type: new
    • Abstract: When a benchmark’s accuracy saturates, it is often retired and replaced with a more challenging version
    • We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues such…
    • We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimen…

• Refusal Lives Downstream of Persona in Chat Models

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26161v1 Announce Type: new.
    • Abstract: In instruction-tuned chat models, linear directions in activation space have been identified for both refusal and persona traits, but the two have been studied as separate mechanisms.
    • We show they interact: a compliant persona gates refusal.
    • In Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, we extract a compliant model-persona direction and a refusal direction and intervene on both.
  • EN Highlights:
    • arXiv:2606.26161v1 Announce Type: new
    • Abstract: Linear directions in activation space have been identified for both refusal and persona traits in instruction-tuned chat models, but the two have been…
    • We show they interact: a compliant persona gates refusal
    • In Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, we extract a compliant model-persona direction and a refusal direction and intervene on both

• AlgoEvolve: LLM-driven Meta-evolution of Algorithmic Trading Programs

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26173v1 Announce Type: new.
    • Abstract: Recent work shows that Large Language Models (LLMs) can act as semantic mutation operators for the evolutionary discovery of programs and proofs.
    • Most current applications have focused on static coding benchmarks.
    • We extend this paradigm to algorithmic trading.
  • EN Highlights:
    • arXiv:2606.26173v1 Announce Type: new
    • Abstract: Recent work shows that Large Language Models (LLMs) can act as semantic mutation operators for the evolutionary discovery of programs and proofs

• Most current applications focus on static coding benchmarks

• We extend this paradigm to algorithmic trading

• Agentic Analysis for Agentic Infrastructure: An LLM-Powered Pipeline for Comparative Governance of DAO and Corporate AI Protocols

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26203v1 Announcement Type: New.
    • Abstract: As AI agent protocols proliferate, the governance structures shaping their interoperability standards remain empirically underexamined.
    • We introduce an LLM-powered comparative pipeline for large-scale governance discourse analysis, integrating automated annotation, neural topic modeling, and multilayer network analysis to study sociotechnical power structures at scale.
    • We validate it on two contrasting standards for agent interoperability: ERC-8004 (permissionless, on-chain) and Google A2A (corporate-led).
  • EN Highlights:
    • arXiv:2606.26203v1 Announce Type: new
    • Abstract: As AI agent protocols proliferate, the governance structures shaping their interoperability standards remain empirically underexamined
    • We introduce an LLM-powered comparative pipeline for large-scale governance discourse analysis, integrating automated annotation, neural topic modeling, and mul…
    • We validate it on two contrasting standards for agent interoperability: ERC-8004 (permissionless, on-chain) and Google A2A (corporate-led)

• Knowledge-augmented Agentic AI for Mental Health Medication Information Seeking

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26205v1 Announcement Type: New.
    • Abstract: Patients increasingly seek medication information online, yet safety knowledge for psychiatric drugs is split between authoritative but abstract regulatory adverse-event records and experiential but unverified patient narratives.
    • Integrating them without conflating evidence and anecdote is especially consequential in psychiatry, where poorly contextualized information can amplify fear, nocebo responses, and non-adherence.
    • Here, we develop a provenance-aware, knowledge graph-based multi-agent framework, unifying 466,525 Reddit posts, 60,782 WebMD reviews, and 20 years of U.S. historical data.
  • EN Highlights:
    • arXiv:2606.26205v1 Announce Type: new
    • Abstract: Patients increasingly seek medication information online, yet safety knowledge for psychiatric drugs is split between regulatory adverse-event records…
    • Integrating them without conflating evidence and anecdote is especially consequential in psychiatry, where poorly contextualised information can amplify fear, n…

• Here we develop a provenance-aware, knowledge-graph-based multi-agent framework unifying 466,525 Reddit posts, 60,782 WebMD reviews, and twenty years of U.S

• Accelerating Skill Assessment in Chess: A Drift-Diffusion-Enhanced Elo Rating System

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26267v1 Announce Type: new.
    • Abstract: Rating systems such as Elo serve as the gold standard for matchmaking in competitive chess.
    • However, they inherently suffer from response lag due to their exclusive reliance on match outcomes, neglecting the granular quality of gameplay.
    • Nevertheless, incorporating move-by-move information into rating adjustments presents a significant challenge given the substantial noise and the vastness of the game’s state space.
  • EN Highlights:
    • arXiv:2606.26267v1 Announce Type: new
    • Abstract: Rating systems such as Elo serve as the gold standard for matchmaking in competitive chess
    • However, they inherently suffer from response lag due to their exclusive reliance on match outcomes, neglecting the granular quality of gameplay
    • Nevertheless, incorporating move-by-move information into rating adjustments presents a significant challenge given the substantial noise and the vastness of th…

• Governing Actions, Not Agents: Institutional Attestation as a Governance Model for Autonomous AI Systems

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26298v1 Announce Type: new.
    • Abstract: Autonomous AI agents may begin to perform consequential, irreversible actions such as clinical prescribing and production software deployment.
    • This paper observes that human institutions govern powerful autonomous actors not by monitoring their reasoning but by requiring independently attested evidence when taking corresponding actions.
    • We formalize this institutional pattern as a computational governance model for AI agent systems.
  • EN Highlights:
    • arXiv:2606.26298v1 Announce Type: new
    • Abstract: Autonomous AI agents may begin to perform consequential, irreversible actions such as clinical prescribing and production software deployment
    • This paper observes that human institutions have governed powerful autonomous actors not by monitoring their reasoning but by requiring independently attested e…
    • We formalise this institutional pattern as a computational governance model for AI agent systems

• COrigami: An AI Pipeline for Co-Designing Flat-Foldable Visually Recognisable Origami

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26299v1 Announce Type: new.

• Abstract: While generative AI has achieved remarkable success in solving problems with verifiable solutions, generating physical art that satisfies both strict geometric constraints and subjective visual aesthetics remains a challenge.

  • This paper presents an approach to tackle these difficulties in the domain of computational origami, a mathematically rigid environment that grounds artistic design on the equations of planar foldability.
  • We present COrigami, an end-to-end AI-driven pipeline that assists the design cycle by generating crease patterns from natural language.
  • EN Highlights:
    • arXiv:2606.26299v1 Announce Type: new
    • Abstract: While generative AI has achieved remarkable success in solving problems with verifiable solutions, generating physical art that satisfies both strict…
    • This paper presents an approach to tackle these difficulties in the domain of computational origami, a mathematically rigid environment that grounds artistic de…
    • We present COrigami, an end-to-end AI-driven pipeline that assists the design cycle by generating crease patterns from natural language

• The Verification Horizon: No Silver Bullet for Coding Agent Rewards

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26300v1 Announce Type: new.
    • Abstract: A classical intuition holds that verifying solutions is easier than generating them.
    • For today’s coding agents, this intuition is being inverted: as foundation models develop stronger reasoning capabilities and engineering tools grow more sophisticated, generating complex candidate solutions is no longer difficult - reliably verifying them has become the harder problem.
    • Every verifier we can build is only a proxy for human intent, never the intent itself.
  • EN Highlights:
    • arXiv:2606.26300v1 Announce Type: new
    • Abstract: A classical intuition holds that verifying a solution is easier than producing one
    • For today’s coding agents, this intuition is being inverted: as foundation models develop stronger reasoning capabilities and engineering harnesses grow more so…
    • Every verifier we can build is only a proxy for human intent, never the intent itself

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

• HierBias: Context-Conditioned Hierarchical Media Bias Detection with Multi-Task Type Classification

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26100v1 Announce Type: new.
    • Abstract: Media bias detection is a critical task for ensuring fair and balanced information dissemination, but existing sentence-level methods classify each sentence independently, ignoring inter-sentence contextual signals naturally utilized by human annotators.
    • We propose \textbf{HierBias}, a hierarchical context-conditioned media bias detector that formally models document context in bias prediction.
    • We introduce \emph{context-conditioned bias probability} and theoretically prove that leveraging document context strictly reduces the Bayesian error of sentence-level classification when inter-sentence mutual information is non-zero.
  • EN Highlights:
    • arXiv:2606.26100v1 Announce Type: new

• Abstract: Media bias detection is a critical task for ensuring fair and balanced information dissemination, yet existing sentence-level approaches classify each…

• We present \textbf{HierBias}, a hierarchical context-conditioned media bias detector that formally models document context in bias prediction

• We introduce the \emph{context-conditioned bias probability} and prove theoretically that leveraging document context strictly reduces the Bayes error of senten…

• Know2Guess: A Contamination-Aware Multi-Zone Benchmark for Knowledge-Boundary Evaluation in Large Language Models

  • Release Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26101v1 Announce Type: new.
    • Abstract: Reliable evaluation of large language models should separate supported answers from unsupported guesses, without conflating them with data contamination, prompt idiosyncrasies, or general refusal behaviors.
    • We present a contamination-aware, multi-zone benchmark for measuring the transition from answerable knowledge to abstention-expected unknowns under frozen build-time labels.
    • The benchmark contains 1,200 items across five domains, explicit abstention expectations, contamination-risk metadata, and dual parsing with an official strict parser and a standardized robust parser.
  • EN Key Points:
    • arXiv:2606.26101v1 Announce Type: new
    • Abstract: Reliable evaluation of large language models should separate supported answering from unsupported guessing without conflating either with data contami…
    • We present a contamination-aware, multi-zone benchmark for measuring the transition from answerable knowledge to abstention-expected unknowns under frozen build…
    • The benchmark contains 1,200 items across five domains, explicit abstention expectations, contamination-risk metadata, and dual parsing with an official strict…

• Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training

  • Release Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26102v1 Announce Type: new.
    • Abstract: Standard post-training processes apply supervised fine-tuning (SFT) and reinforcement learning (RL) to make language models helpful, but these processes can unintentionally degrade values instilled during pre-training.
    • We study whether the domain of post-training data differentially impacts the retention of animal compassion values in a Llama 3.1 8B model mid-trained on compassion-oriented synthetic data, using SFT (helpfulness via Dolly-15k vs. coding via Magicoder-110K) and GRPO (with helpfulness via RLHFlow).
  • EN Key Points:
    • arXiv:2606.26102v1 Announce Type: new

• Abstract: Standard post-training pipelines apply supervised fine-tuning (SFT) and reinforcement learning (RL) to make language models helpful, but these process…

• We investigate whether the domain of post-training data differentially affects the retention of animal compassion values in a Llama 3.1 8B model mid-trained on…

• coding via Magicoder-110K) and GRPO (helpfulness via RLHFlow vs

• Investigating LLM’s Problem Solving Capability – a Study on Statics Questions

  • Posted: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26103v1 Announcement Type: new.
    • Abstract: Large Language Models (LLMs) have rapidly influenced many aspects of society, particularly education, given their ability to complete assignments and exams across a wide range of disciplines.
    • Although prior studies have examined the educational impact of LLMs, most existing work relies on public or open-source problem datasets and lacks topic-specific analysis.
    • In engineering education, especially within mechanical engineering, systematic investigations into LLM performance on specific problem types remain limited.
  • EN Key Points:
    • arXiv:2606.26103v1 Announce Type: new
    • Abstract: Large Language Models (LLMs) have rapidly influenced many aspects of society, particularly education, due to their demonstrated ability to complete as…
    • Although prior studies have examined the educational impact of LLMs, much of the existing work relies on public or open problem datasets and lacks topic-specifi…
    • In engineering education, especially within mechanical engineering, systematic investigations of LLM performance on specific problem types remain limited

• Assert, don’t describe: Linguistic features that shift LLM reasoning about animal welfare

  • Posted: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26104v1 Announcement Type: new.
    • Abstract: Animal welfare advocates write a large volume of articles, and increasingly these texts are used to train the language models that millions of people then ask about animal welfare.
    • Using stance-contrast probes with lexical matching on a held-out animal welfare benchmark, we measure how each of ten linguistic features, when used as fine-tuning data, alters the preference of a Llama-3.2-1B model for pro-animal welfare reasoning.
    • Eight of the ten features produce statistically significant changes.
  • EN Key Points:
    • arXiv:2606.26104v1 Announce Type: new
    • Abstract: Animal-welfare advocates produce a lot of writing, and increasingly that writing trains the language models that millions of people then ask about ani…

• Using vocabulary-matched stance-contrast probes on a held-out animal-welfare benchmark, we measure how each of ten linguistic features changes Llama-3.2-1B’s pr…

• Eight of the ten features produce statistically significant shifts

• Context Recycling for Long-Horizon LLM Inference

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26105v1 Announce Type: new.
    • Abstract: Large language models (LLMs) exhibit strong capabilities in short-context reasoning, but their performance degrades over long conversational horizons due to context window limitations and inefficient token usage.
    • We introduce ContextForge, a context recycling system that maintains task-relevant information across turns by combining structured query generation, external memory retrieval, and controlled synthesis.
    • This system enables the efficient reuse of prior computations without relying on full context replay, reducing token overhead while preserving answer quality.
  • EN Key Points:
    • arXiv:2606.26105v1 Announce Type: new
    • Abstract: Large language models (LLMs) exhibit strong capabilities in short-context reasoning but degrade in performance over long conversational horizons due t…
    • We introduce ContextForge, a system for context recycling that maintains task-relevant information across turns by combining structured query generation, extern…
    • The system enables efficient reuse of prior computation without relying on full context replay, reducing token overhead while preserving answer quality

• Reducing Conversational Escalation in Large Language Model Dialogue with Nonviolent Communication Constraints

  • Published: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26106v1 Announce Type: new.
    • Abstract: Large language models (LLMs) are increasingly used in emotionally charged situations involving interpersonal conflict, frustration, and distress.
    • While prior safety research has focused on preventing explicit harms such as toxic or policy-violating content, less attention has been paid to conversational behaviors that may inadvertently escalate conflict.
    • In this paper, we investigate whether LLMs can be guided toward more de-escalatory conversational behaviors through lightweight, prompt-level constraints derived from Nonviolent Communication (NVC).
  • EN Key Points:
    • arXiv:2606.26106v1 Announce Type: new
    • Abstract: Large language models (LLMs) are increasingly used in emotionally charged situations involving interpersonal conflict, frustration, and distress
    • While prior safety research has focused on preventing explicit harms such as toxic or policy-violating content, less attention has been paid to conversational b…

• In this paper, we investigate whether LLMs can be guided toward more de-escalating dialogue behavior through lightweight prompt-level constraints derived from N…

• Low Resource Multimodal Translation of Nepali Spoken Words into Emotion-Conditioned Sign Language Avatars

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26107v1 Announcement Type: New.
    • Abstract: Sign language communication systems that integrate emotional expression remain underexplored, especially for low-resource languages.
    • This pilot study presents NEST-V1 (Nepali Emotion and Speech Transformer - Version 1), a proof-of-concept multimodal framework that demonstrates the feasibility of generating emotion-conditioned Nepali sign language avatars from spoken input.
    • As a preliminary investigation, we focus on four common Nepali words (“thank you,” “hello,” “house,” “me”) across three emotional states (happy, neutral, sad) to validate our core technical approach.
  • EN Highlights:
    • arXiv:2606.26107v1 Announce Type: new
    • Abstract: Sign language communication systems, that integrate emotional expression remain underexplored, particularly for low-resource languages
    • This pilot study presents NEST-V1 (Nepali Emotion and Speech Transformer - Version 1), a proof-of-concept multimodal framework that demonstrates the feasibility…
    • As a preliminary investigation, we focus on four common Nepali words (“thank you”, “hello”, “house”, “me”) across three emotional states (happy, neutral, sad) t…

• Where Larger Models Excel: The Primacy of Constraint-Guided Reasoning

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26108v1 Announcement Type: New.
    • Abstract: Larger language models consistently outperform smaller language models on reasoning benchmarks, but the reasoning differences behind this gap remain underexplored.
    • In benchmarks for mathematics, physics, chemistry, and programming, we observe stable performance gaps: on average across datasets, Qwen3-32B outperforms Qwen3-8B by 6.43%, while GPT-OSS-120B outperforms GPT-OSS-20B by 7.38%.
    • To investigate the reasoning differences behind these gains, we developed AdvCluster, an automated framework that identifies problems where larger models show a stable advantage, extracts fine-grained descriptions of the advantage from paired reasoning trajectories generated by the larger and smaller models, organizes them via semantic clustering, and performs quantitative evaluation and selection under the guidance of a reviewer model.
  • EN Highlights:
    • arXiv:2606.26108v1 Announce Type: new
    • Abstract: Larger language models consistently outperform smaller ones on reasoning benchmarks, yet the reasoning differences underlying this gap remain underexp…
    • Across benchmarks in mathematics, physics, chemistry, and programming, we observe stable performance gaps: averaged over datasets, Qwen3-32B outperforms Qwen3-8…

• To study the reasoning differences behind these gains, we develop AdvCluster, an automated framework that identifies questions where the larger model shows a st…

• From Lexicon to AI: A Structured-Data Pipeline for Specialized Conversational Systems in Low-Resource Languages

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26112v1 Announcement Type: new.
    • Abstract: Low-resource languages face a critical challenge in AI development: creating specialized conversational systems without access to massive training corpora.
    • We propose a systematic methodology for transforming structured linguistic resources into specialized AI systems, demonstrating that expert-curated lexical databases can serve as an effective foundation for conversational AI development.
    • Our approach converts Hindi WordNet into 1.25 million diverse instruction-response pairs, fine-tuning a 12B-parameter language model using resource-efficient LoRA with 4-bit quantization.
  • EN Highlights:
    • arXiv:2606.26112v1 Announce Type: new
    • Abstract: Low-resource languages face a critical challenge in AI development: creating specialized conversational systems without access to massive training cor…
    • We present a systematic methodology for transforming structured linguistic resources into specialized AI systems, demonstrating that expert-curated lexical data…
    • Our approach converts Hindi WordNet into 1.25 million diverse instruction-response pairs, fine-tunes a 12B-parameter language model using resource-efficient LoR…

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

• Physics-guided Convolutional Neural Network for Domain Growth Prediction in Systems with Conserved Kinetics

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26128v1 Announcement Type: new.
    • Abstract: The spatiotemporal evolution of many physical, chemical, and biological systems is described by nonlinear partial differential equations (PDEs).
    • Recently, deep neural network-based surrogate models have gained increasing attention as effective alternatives to computationally expensive traditional numerical solvers.
    • In this work, we propose an attention-based, physics-guided convolutional neural network as a surrogate model to learn the microstructural evolution of such systems.
  • EN Highlights:
    • arXiv:2606.26128v1 Announce Type: new
    • Abstract: The spatiotemporal evolution of many physical, chemical, and biological systems is described by nonlinear partial differential equations (PDEs)
    • Recently, deep neural network-based surrogate models have gained increasing interest as efficient alternatives to computationally expensive traditional numerica…

• In this work, we propose an attention-based, physics-guided convolutional neural network as a surrogate model to learn the microstructural evolution of such sys…

• \chisao{}: A GPU-Native Parallel Optimizer for Multimodal Black-Box Functions via Convergence-Anticonvergence Oscillation

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26164v1 Announce Type: new.
    • Abstract: Finding all modes of a multimodal black-box function is a fundamental challenge in optimization, Bayesian inference, and scientific computing.
    • Existing methods—basin-hopping, CMA-ES, multi-start gradient descent—operate sequentially and cannot leverage the massive parallelism of modern GPU hardware.
    • We introduce \chisao{} (\textbf{C}onvergence-\textbf{H}alt-\textbf{I}nvert-\textbf{S}tick-\textbf{A}nd-\textbf{O}scillate), a GPU-native population optimizer that simultaneously runs on an entire batch of samples and utilizes intentional convergence-anticonvergence oscillation cycles to escape local traps while freezing confirmed modes.
  • EN Key Points:
    • arXiv:2606.26164v1 Announce Type: new
    • Abstract: Finding all modes of a multimodal black-box function is a fundamental challenge in optimization, Bayesian inference, and scientific computing
    • Existing approaches – basin-hopping, CMA-ES, multistart gradient descent – operate sequentially and cannot exploit the massive parallelism of modern GPU hardw…
    • We introduce \chisao{} (\textbf{C}onvergence-\textbf{H}alt-\textbf{I}nvert-\textbf{S}tick-\textbf{A}nd-\textbf{O}scillate), a GPU-native population optimizer th…

• Implementation of reinforcement learning in chemical reaction networks: application to phototaxis as curiosity-driven exploration

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26168v1 Announce Type: new.
    • Abstract: Living systems use noisy and incomplete sensory signals to navigate their environment.
    • In unicellular algae, phototaxis is often modeled as a mechanistic run-and-tumble process driven by stimulus-response rules.
    • However, such descriptions overlook how organisms actively sample their environment to reduce sensory ambiguity.
  • EN Key Points:
    • arXiv:2606.26168v1 Announce Type: new
    • Abstract: Living systems navigate environments using noisy and incomplete sensory signals
    • In unicellular algae, phototaxis is often modeled as a mechanistic run–tumble process driven by stimulus–response rules
    • However, such descriptions overlook how organisms actively sample their environment to reduce sensory ambiguity

• Neural Architecture Search for Generative Adversarial Networks: A Comprehensive Review and Critical Analysis

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26169v1 Announcement Type: new.
    • Abstract: Neural Architecture Search (NAS) has become a key technology for optimizing the design of Generative Adversarial Networks (GANs), automatically searching for effective architectures while addressing the challenges inherent in manual design.
    • This paper provides a comprehensive review of NAS methods applied to GANs, classifying and comparing various approaches based on criteria such as search strategy, evaluation metrics, and performance results.
    • The review highlights the advantages of NAS in improving GAN performance, stability, and efficiency, while also identifying limitations and areas for future research.
  • EN Key Points:
    • arXiv:2606.26169v1 Announce Type: new
    • Abstract: Neural Architecture Search (NAS) has emerged as a pivotal technique in optimizing the design of Generative Adversarial Networks (GANs), automating the…
    • This paper provides a comprehensive review of NAS methods applied to GANs, categorizing and comparing various approaches based on criteria such as search strate…
    • The review highlights the benefits of NAS in improving GAN performance, stability, and efficiency, while also identifying limitations and areas for future resea…

• KG-TRACE: A Neuro-Symbolic Framework for Mechanistic Grounding in Antimicrobial Resistance Prediction

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26179v1 Announcement Type: new.
    • Abstract: Although WGS-based AMR prediction has achieved high accuracy, existing models lack a mechanism for grounding neural attributions in established biological pathways.
    • We propose KG-TRACE, a novel neuro-symbolic framework that integrates the WHO mutation knowledge graph (KG) as a structured biological constraint on a neural genomic model.
    • Unlike existing methods that learn statistical patterns in isolation, KG-TRACE fuses genomic features and RotatE-based KG embeddings through a learned epistemic trust gate, dynamically weighting neural evidence based on symbolic biological knowledge.
  • EN Key Points:
    • arXiv:2606.26179v1 Announce Type: new
    • Abstract: While WGS-based AMR prediction has reached high accuracy, existing models lack a mechanism to ground neural attributions in established biological pat…
    • We present KG-TRACE, a novel neuro-symbolic framework that integrates the WHO mutation knowledge graph (KG) as a structured biological constraint on a neural ge…
    • Unlike existing methods that learn statistical patterns in isolation, KG-TRACE fuses genomic features and RotatE-based KG embeddings through a learned epistemic…

• Necessary but Not Sufficient: Temperature Control and Reproducibility in LLM-as-Judge Safety Evaluations

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Summary: - arXiv:2606.26185v1 Announcement Type: New.
    • Summary: LLM-as-judge (“grader”) components are now a standard configuration in evaluation tools, including safety evaluations, where pass/fail verdicts may influence downstream deployment decisions.
    • A widespread assumption is that setting the grader’s sampling temperature to 0 makes the grading deterministic.
    • We test this assumption against a real safety evaluation codebase (Japan AISI’s open-source aisev) and show that it fails on two levels.
  • EN Highlights:
    • arXiv:2606.26185v1 Announce Type: new
    • Abstract: LLM-as-judge (“grader”) components are now standard in evaluation harnesses, including safety evaluations where a pass/fail verdict may gate downstrea…
    • A widespread assumption is that setting the grader’s sampling temperature to 0 makes grading deterministic
    • We test this assumption against a real safety-evaluation codebase (Japan AISI’s open-source aisev) and show it fails on two levels

• Clue-Guided Money Laundering Group Discovery

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Summary: - arXiv:2606.26189v1 Announcement Type: New.
    • Summary: Money Laundering Group Discovery (MLGD) aims to identify hidden criminal groups and recover their complete structures in large-scale financial networks.
    • Existing graph anomaly detection methods mainly produce node-level risk alerts, while global group discovery methods passively search for suspicious groups across the entire network.
    • Both are inconsistent with real Anti-Money Laundering (AML) investigations, where analysts typically start with a specific clue and gradually expand the scope of their investigation to track down the responsible parties.
  • EN Highlights:
    • arXiv:2606.26189v1 Announce Type: new
    • Abstract: Money Laundering Group Discovery (MLGD) aims to identify hidden criminal groups and recover their complete structures in large-scale financial network…
    • Existing graph anomaly detection methods mainly produce node-level risk alerts, while global group discovery methods passively search for suspicious groups over…
    • Both are mismatched with real Anti-money-laundering (AML) investigations, where analysts usually start from a concrete clue and gradually expand the investigati…

• Federated Hash Projected Latent Factor Learning

  • Publication Time: 2026-06-26 12:00 Beijing Time
  • Summary: - arXiv:2606.26192v1 Announcement Type: New.
    • Summary: Hash Learning (HL) is an effective representation learning method that can map real-valued data into a compact binary representation.
    • Traditional HL methods typically require users to upload personal data to a central server, which is incompatible with increasingly strict data security regulations.
    • Federated Learning (FL) provides a decentralized paradigm for learning a globally optimal model without centralizing private data.

• EN Key Points:

  • arXiv:2606.26192v1 Announce Type: new
  • Abstract: Hash Learning (HL) is an efficient representation learning approach that maps real-valued data into compact binary representations
  • Traditional HL methods typically require users to upload personal data to a central server, which is incompatible with increasingly stringent data security regu…
  • Federated Learning (FL) provides a decentralized paradigm for learning globally optimal models without centralizing private data

• Statistical and Structural Approaches to Algorithmic Fairness

  • Publish Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26200v1 Announce Type: new.
    • Abstract: Modern machine learning systems have outgrown their origins as isolated predictive constructs, evolving into complex socio-technical architectures that actively regulate human opportunities.
    • As algorithms increasingly determine access to economic and social opportunities, it has become widely recognized that these systems are deeply embedded with the structural inequalities and biases of their environment.
    • There is a growing recognition that models optimized for predictive accuracy can systematically disadvantage marginalized groups, and thus the field of algorithmic fairness emerged in response to this awareness.
  • EN Key Points:
    • arXiv:2606.26200v1 Announce Type: new
    • Abstract: Modern machine learning systems have outgrown their origins as isolated predictive constructs, evolving into complex socio-technical architectures tha…
    • As algorithms increasingly determine access to economic and social opportunities, it has become widely recognized that these systems are deeply embedded with th…
    • The field of algorithmic fairness emerged in response to the growing recognition that models optimized for predictive accuracy can systematically disadvantage m…

• Topology-Informed Neural Networks for Flood Detection in Optical and Synthetic Aperture Radar Imagery

  • Publish Time: 2026-06-26 12:00 Beijing Time
  • Abstract: - arXiv:2606.26204v1 Announce Type: new.
    • Abstract: Floods frequently impact regions around the world.
    • Rapid and accurate flood detection is crucial for emergency response and timely mitigation of human and economic loss.
    • The expanding availability of satellite data and advancements in AI have enhanced the monitoring of environmental hazards, but many flood events remain difficult to detect due to cloud cover obscuring optical satellite imagery.
  • EN Key Points:
    • arXiv:2606.26204v1 Announce Type: new
    • Abstract: Floods frequently impact regions around the world
    • Rapid and accurate flood detection is crucial for emergency response and timely mitigation of human and economic loss

• The expanding availability of satellite data and advances in artificial intelligence have enhanced monitoring of environmental hazards, but many flood events re…