2026-06-12 AI Daily | Fable 5 Hits Cost Warning Threshold, AI Agents Are Fostering a New “Harness” Engineering Discipline Link to heading

The buzz around Claude Fable 5 validates the feasibility of long-context reasoning agents, but its daily consumption has already hit a threshold equivalent to thousands of dollars, forcing the industry to confront tokenomics. Meanwhile, on-device models are transitioning from “usable” to “practical,” and DeepSeek’s creation of an “Agent Harness Researcher” position signals that the engineering system connecting foundational models to autonomous behavior is emerging as a new, independent discipline.

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

Today, the AI frontier reveals two clear, intersecting trends. First, the deep evolution of Agents from “execution” to “decision-making.” Several papers worth a close read have emerged on arXiv: INFRAMIND moves beyond simple model routing to directly sense the runtime state of the underlying GPU cluster for multi-agent scheduling. Self-Gated Clarification allows agents to learn to proactively decide “when to ask” at crucial reasoning junctures, directly tackling the fragility of current long-cycle agent implementations. For teams building research or financial agents, SciConBench and MoCA-Agent provide new benchmarks and architectures, focusing on synthesizing scientific conclusions and financial numerical reasoning, respectively.

Second, new perspectives on the internal structure of large model alignment and behavioral prediction. Dual-Stance Evaluation reveals that sycophancy and factual consistency occupy geometrically distinct subspaces, making precise steering possible. A position paper directly proposes treating behavior prediction itself as a learnable task, bypassing the chain of explanation. In addition, a high-quality interview is also worth a listen—Ben Bajarin provides a deep dive into Apple’s latest strategies in AI and computing, offering a very restrained industry benchmark for on-device intelligence.

🌐 AI Hot Topics on X Link to heading

Topic 1: Anthropic’s Claude Code Event Packs Tokyo with Developers Link to heading

• Category: AI · News
• Summary: Trending since: 17 hours ago, Related posts: 223
• What happened: Anthropic’s Claude Code developer event in Tokyo was packed, with a large number of local developers in attendance.
• Why it’s important: This signals Anthropic’s accelerated expansion of its developer ecosystem in Asia, strengthening its global competitive position against rivals like OpenAI and GitHub Copilot in the AI coding assistant space, and exploring the market potential for enterprise-level AI tools in Japan.
• Discussion summary: The discussion on X focuses on Claude Code’s actual coding performance compared to GitHub Copilot, the practicality of Anthropic’s promotion strategy in Asia, and a debate among some developers about whether the event’s content was more focused on technical substance or brand marketing.

Topic 2: OpenAI Hires Cybersecurity Leaders to Counter AI Risks Link to heading

• Category: AI · News
• Summary: Trending since: 23 hours ago, Related posts: 270
• What happened: OpenAI has hired cybersecurity leaders to strengthen its own defenses and address potential security threats posed by AI.
• Why it’s important: This move indicates that leading AI companies are investing resources in governance and defense to proactively address new cybersecurity risks arising from the potential misuse of AI technology, serving as a bellwether for industry security practices and standards.
• Discussion summary: Discussions on X are centered on the background and past experience of the new hires, whether this move marks a substantive step in OpenAI’s commitment to safety, and its potential impact on AI industry security standards. Some opinions also question if this action is related to recent internal safety controversies.

Topic 3: OpenClaw Developer Open-Sources AI Repo Maintainer Skills Link to heading

• Category: AI · News
• Summary: Trending since: 9 hours ago, Related posts: 488
• What happened: An OpenClaw developer has open-sourced a set of “AI Repo Maintainer Skills,” encapsulating common maintenance tasks like code review, issue triage, and documentation updates into reusable, automated capabilities.
• Why it’s important: This initiative aims to lower the maintenance costs and barriers for open-source AI projects by allowing agents to replace or assist humans in handling repetitive maintenance work. It has the potential to accelerate the democratization of AI toolchains and improve community collaboration efficiency.
• Discussion summary: The community’s focus is on: 1) Whether automated maintenance can truly guarantee code quality and security, or if it might introduce subtle errors; 2) Whether the role of human maintainers will be marginalized; 3) How skill templates can accommodate the different maintenance styles of various projects, and whether this could lead to a homogenization of maintenance practices.

Topic 4: Peter Steinberger Open-Sources AI Skills for Autonomous Repo Maintenance Link to heading

• Category: AI · News
• Summary: Trending since: 10 hours ago, Related posts: 609
• What it is: Developer Peter Steinberger has open-sourced a set of AI skills for autonomous code repository maintenance.
• Why it matters: This move lowers the barrier to entry for developing autonomous software maintenance agents, advancing AI from code generation to long-term, reliable, repository-level automated management. It helps validate the reliability of large models in real-world, continuous engineering tasks.
• Discussion overview: The main focuses include: the reliability boundaries of autonomous bug fixing and PR submission, how to prevent automation from introducing new problems, and effective ways to integrate open-source solutions with existing CI/CD and code review processes.

Topic 5: Recursive AI Tops Benchmarks in Automated Research Breakthrough Link to heading

• Category: AI · News
• Overview: Trending time: 13 hours ago, Related posts: 540
• What it is: A recursive AI system has achieved the highest scores on automated scientific research benchmarks, marking a breakthrough.
• Why it matters: This demonstrates AI’s ability to conduct scientific research autonomously and accelerate discovery through recursive self-improvement. It could reshape the paradigm of scientific R&D, while also raising deep concerns about the safety and control of recursive self-improvement.
• Discussion overview: Discussions on the X platform center on whether this achievement represents a genuine breakthrough in autonomous research or just benchmark chasing; the potential risks of loss of control and alignment challenges posed by recursive self-improvement; and whether the related models are open-source and the results reproducible. Optimists believe this ushers in an AI-driven scientific revolution, while pessimists stress the need for strengthened safety regulations in advance.

Topic 6: Tesla Deploys FSD Supervised v14.3.4 with Smart Summon for Cybertruck Link to heading

• Category: AI · News
• Overview: Trending time: , Related posts: 4300
• What it is: Tesla has started rolling out FSD Supervised v14.3.4 to the Cybertruck, adding the Smart Summon feature.
• Why it matters: This indicates Tesla is further unifying the experience across its vehicle lineup in its fully autonomous driving software iterations. It also brings the controversial Cybertruck into the core L2 assisted driving ecosystem, which will help collect road data from this unique platform and advance the generalization capabilities of its end-to-end model.
• Discussion overview: Users are primarily focused on the system’s performance on unpaved roads and in adverse weather. Some question the impact of the Cybertruck’s unusual design on camera visibility. There are also heated discussions about the practicality of Smart Summon in tight parking spaces, its improvements over previous versions, and whether this version still requires continuous driver attention monitoring.

AI Public Opinion Summary on X Today Link to heading

The main narrative in today’s public opinion clearly points to the evolution of AI systems from auxiliary tools to long-term, autonomous executors of complex tasks. This trend is sweeping across fields like software maintenance, scientific research, and physical driving. An industry consensus is emerging as leading companies actively position themselves to capture ecological niches—Anthropic is focused on expanding its Asian developer community, OpenAI is strengthening its safety governance, and Tesla is extending its end-to-end driving model to controversial vehicle models. All parties agree that open-sourcing and automation are key to lowering barriers and accelerating iteration. Disagreements, however, center on the true reliability of autonomous capabilities. The community is hotly debating whether automated maintenance will introduce hidden defects, whether recursive AI is a genuine research breakthrough or just benchmark chasing, and the impact of the Cybertruck’s unique design on driving perception. Potential risks are becoming increasingly prominent: unattended agents could introduce imperceptible errors, the loss of control and alignment challenges of recursive self-improvement are causing deep concern, and the new types of cyber threats spawned by AI misuse make upgrading security defenses an urgent industry-wide imperative.

💡 Influencer Insights Link to heading

AI Daily: Claude Fable 5 Ignites a New Paradigm in Agent Development, Amidst On-Device Models and Cost Anxiety Link to heading

I. Today’s Core Hotspot: Claude Fable 5 and the Agent Development Paradigm Link to heading

1.1 Fable 5: A Leap in Capability Amidst Cost Controversy Link to heading

Claude Fable 5 is the absolute focus today, with multiple influencers conducting intensive tests and providing in-depth reviews:

1.2 Evolution of Agent Development Paradigms Link to heading

“Goal Instruction” becomes the new standard for Codex/Claude Code:

• @vista8 developed “Qiaomu Goal Meta Skill” (npx skills add joeseesun/qiaomu-goal-meta-skill), transforming single-sentence requirements into executable goals, supporting a long-duration task mode of “execute before bed, harvest the next day”
• @dotey practiced the /goal instruction, significantly improving long-task stability, with “no need to continue”

Extreme Case of Fable 5: @trq212 demonstrated a video production workflow entirely generated by AI coding—Whisper transcription → Subagent clip selection → FFmpeg rough cut → hand-written LUTs color grading → Remotion animation components → Figma MCP collaboration, with no traditional non-linear editing software intervention throughout.

II. Edge-Side Models: Performance Breakthroughs and Ecosystem Maturity Link to heading

2.1 Actual Progress Link to heading

@zhixianio’s “ascetic practice” verified that edge-side models are already productive:

• Configuration: Qwen3.6-35B-A3B-oQ6-fp16-mtp / oMLX / Native MTP / 128K CTX
• Conclusion: “Response speed is faster than remote LLMs, intelligence is online, and native multimodal is even more satisfying than DSV4 Pro”

Google Gemma 4 Series continues to iterate:

• The 12B multimodal model on M5Max 128G recognized English/Japanese “instantly,” but Chinese was “nonsensical”
• QAT (Quantization-Aware Training) new approach: assumes quantization during the training phase, “thereby improving training effectiveness”

2.2 Toolchain Improvement Link to heading

III. Unique Perspectives and Industry Foresight Link to heading

3.1 Cost Anxiety and Business Model Restructuring Link to heading

“AI is more expensive than employees” becomes the new reality:

• @ruanyf calculated: OpenClaw founder’s monthly consumption of 603 billion Tokens, equivalent to $1.3 million; even if using domestic open-source models (1/30-1/50 the price), the annual cost still reaches 2-3 million RMB
• @lijigang proposed the “real metrics” theory: Token consumption is a false metric; “whether the problem is better solved” is the real metric

New business model exploration:

• @oran_ge (via @lijigang): “First version free, subsequent updates charged”—because the first version of AI Coding is the simplest, and maintenance is the most labor-intensive
• @dotey quoted: OpenDoor laid off 200+ offshore team members in India, replacing them with a “smaller, AI-native team in the US”

3.2 Rethinking the Essence of Software Engineering Link to heading

@dotey’s core assertion: “AI has not redefined software engineering; AI has amplified the importance of software engineering”

@Pluvio9yte’s practical verification:

“The best practice for Vibe Coding is not Requirement First or Code First, but Contract First. Without well-defined contracts, everything else is empty talk.”

His development framework, based on a secondary development of OpenSpec, aims to “externalize easily drifting context into contracts, providing stable references for both humans and AI.”

3.3 DeepSeek’s “Harness” Strategy Link to heading

@dotey revealed DeepSeek is recruiting “Agent Harness Researchers”—the first explicit recruitment for this position worldwide:

“Model + Harness = Agent. All work other than the model itself belongs to Harness: context management, long-term memory, Subagent & Multi-Agent, self-evolving Agent…”

Defining “Harness Engineering” as a new discipline, requiring candidates to be “heavy Agent users, full-stack developers, driving research from 0 to 1.”

IV. Recommended Tools and Resources Link to heading

4.1 Development Tools Link to heading

4.2 Content Creation Link to heading

4.3 Infrastructure Link to heading

• OfoxAI: A relay station balancing “stability + discounts,” offering a 15% discount on GPT-5.5/5.4 mini (@AI_Jasonyu)
• Vercel: “The fastest way to launch a website,” deploy with the Codex plugin in just a few sentences (@Pluvio9yte)

V. Key Trend Summary Link to heading

┌─────────────────────────────────────────┐
│  今日核心矛盾：能力跃升 ↑  vs  成本失控 ↑  │
├─────────────────────────────────────────┤
│  • Fable 5 验证"长思考"Agent 的可行性    │
│  • 端侧模型从"能用"走向"好用"            │
│  • "Harness Engineering"成为新工程学科  │
│  • Token 经济学倒逼商业模式创新           │
│  • Contract First 取代 Vibe Coding 随意性 │
└─────────────────────────────────────────┘

Focus for Tomorrow: User retention strategies for Fable 5 before the June 22 subscription deadline, actual test results for Gemma 4 QAT, and progress on the DeepSeek Harness team formation.

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

Time window: Last 3 days; covers 22 sources; 35 updates in total

Stratechery by Ben Thompson (A_full) Link to heading

• An Interview with Ben Bajarin About Apple, AI, and Compute
  • Published: 2026-06-11 18:00 Beijing Time
  • Summary: - An interview with Ben Bajarin about WWDC and the current state of the AI compute industry.
    • $15/month* or *$150/year.
    • Substantive analysis of the day’s news via three weekly emails or a podcast.
    • Strategy Interviews.
    • Interviews with leading public company CEOs, private company founders, and discussions with fellow analysts.
  • EN Key Points:
    • An interview with Ben Bajarin about WWDC and the status of the AI compute industry.

OpenAI Blog (A_full) Link to heading

• OpenAI to acquire Ona

  • Published: 2026-06-11 08:00 Beijing Time
  • Summary: - Over 5 million people use Codex weekly to research, analyze, build, and automate their work, a 400% increase from earlier this year.
    • Originally a tool for software developers, Codex now helps a broader audience complete complex work from initial request to final result.
    • As Codex becomes more powerful, its most valuable work will unfold over hours or days, not minutes.
    • We believe people should be able to delegate more ambitious work without being tied to the machine where the work began.
    • Work should continue after the initial session, and Codex allows people to stay connected from anywhere to check progress, provide direction, make decisions, and review results.
  • EN Key Points:
    • OpenAI plans to acquire Ona to expand Codex with secure, persistent cloud environments, enabling long-running AI agents across enterprise workflows.

• Supporting Europe’s work in ensuring a trustworthy AI ecosystem

  • Published: 2026-06-11 08:00 Beijing Time
  • Summary: - People are using artificial intelligence to create and edit content in new ways.
    • As these tools become more powerful and widely used, people should understand the context of the content they see online.

• This code of practice is a significant step towards implementing the EU AI Act and building a more transparent digital ecosystem.

• Our support is built on years of internal research, product development, and collaboration with the broader ecosystem to enhance the provenance of AI-generated content.

• Drawing on years of expertise, we, along with hundreds of other stakeholders, contributed to the development of this code to help ensure the establishment of a trustworthy AI ecosystem.

• EN Highlights:

  • OpenAI supports the EU Code of Practice on AI content transparency, advancing provenance standards and tools to help people understand AI-generated content.

• How an astrophysicist uses Codex to help simulate black holes

  • Published: 2026-06-11 08:00 Beijing Time
  • Summary: - Learn how astrophysicist Chi-kwan Chan uses Codex to build black hole simulations, helping scientists study extreme physics and test Einstein’s theory of general relativity.
    • Learn how astrophysicist Chi-kwan Chan uses Codex to build black hole simulations, helping scientists study extreme physics and test Einstein’s generative theory.
    • How an astrophysicist uses Codex to help simulate black holes.
  • EN Highlights:
    • Discover how astrophysicist Chi-kwan Chan uses Codex to build black hole simulations, helping scientists study extreme physics and test Einstein’s theory of gen…

• BBVA puts AI at the core of banking with OpenAI

  • Published: 2026-06-11 08:00 Beijing Time
  • Summary: - Learn how BBVA scaled ChatGPT Enterprise to 100,000 employees and partnered with OpenAI to accelerate AI-powered banking transformation worldwide.
    • This article from the OpenAI blog explains how BBVA is putting AI at the core of banking and shaping the broader AI and infrastructure landscape through OpenAI.
    • Following BBVA’s move to place AI at the core of banking through OpenAI, this also brings practical implications for founders, operators, and investors.
  • EN Highlights:
    • Learn how BBVA scaled ChatGPT Enterprise to 100,000 employees and partnered with OpenAI to accelerate AI-powered banking transformation worldwide.

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

• From Explicit Elements to Implicit Intent: A Predefined Library for Auditable Behavioral Inference

  • Published: 2026-06-11 12:00 Beijing Time
  • Summary: - arXiv:2606.11207v1 Announcement Type: New.
    • Abstract: We introduce SemantiClean, a modular framework for extracting structured semantic signals from e-commerce session data and driving pluggable inference targets—including purchase intent, customer segmentation, and product affinity—through a shared element library.
    • Unlike traditional end-to-end predictors optimized solely for accuracy, SemantiClean prioritizes auditability, structural governance, and sigma=0 reproducibility, explicitly trading marginal predictive gains for element-level transparency and defensible decision trajectories.
    • Grounded in the Online Shopper Purchase Intent (OSPI) dataset, the framework organizes 24 behavioral elements into a four-tier architecture (Functional, Interaction, Systemic, Contextual) and enforces signal quality through three anti-inflationary mechanisms: RedundancyGroup contribution caps, TieredPenaltyCalculator deviation penalties, and AdaptiveConstraintMode cold-start protections. This report introduces the LLM-Integrated Semantic Inference Engine, a fully-realized, two-stage LLM-driven reasoning architecture that leverages complete element metadata at inference time.
  • EN Highlights:

• arXiv:2606.11207v1 Announce Type: new

• Abstract: We present SemantiClean, a modular framework for extracting structured semantic signals from e-commerce session data and driving pluggable inference t…

• Unlike conventional end-to-end predictors that optimise solely for accuracy, SemantiClean prioritises auditability, structural governance, and sigma=0 reproduci…

• Built upon the Online Shoppers Purchasing Intention (OSPI) dataset, the framework organises twenty-four behavioural elements into a four-layer architecture (Fun…

• Position: Hippocampal Explicit Memory Is the Cornerstone for AGI

  • Published: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11245v1 Announce Type: new.
    • Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, raising expectations for Artificial General Intelligence (AGI).
    • This position paper argues that integrating explicit memory is the cornerstone for advancing LLMs toward AGI.
    • The key reason is that the underlying learning mechanism of LLMs is highly analogous to human implicit memory.
  • EN Key Points:
    • arXiv:2606.11245v1 Announce Type: new
    • Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, raising expectations for Artificial General Intelligence…
    • This position paper argues that integrating explicit memory is the cornerstone for advancing LLMs toward AGI
    • The key reason is that the underlying learning mechanism of LLMs is highly analogous to human implicit memory

• Can AI Agents Synthesize Scientific Conclusions?

  • Published: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11337v1 Announce Type: new.
    • Abstract: Scientific AI agents are increasingly retrieving evidence, reasoning across sources, and synthesizing conclusions for subsequent decision-making.
    • However, their ability to do so in high-stakes domains like health remains unclear.
    • We introduce SciConBench, a large-scale real-time benchmark containing 9.11K questions and expert-written conclusions from systematic reviews, to evaluate open-domain scientific conclusion synthesis.
  • EN Key Points:
    • arXiv:2606.11337v1 Announce Type: new
    • Abstract: Scientific AI agents increasingly retrieve evidence, reason across sources, and synthesize conclusions used in consequential decisions
    • Yet, their ability to do so in high-stakes domains such as health remains unclear

• We introduce SciConBench, a large-scale live benchmark of 9.11K questions and expert-written conclusions from systematic reviews to evaluate open-domain scienti…

• Knowing When to Ask: Self-Gated Clarification for Hierarchical Language Agents

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11349v1 Announcement Type: New.
    • Abstract: In hierarchical reasoning, failures often originate at intermediate decision points, where the agent commits to a wrong branch without realizing it lacks critical information.
    • Rather than treating clarification as an external uncertainty trigger, we propose ACTION-RATING, a formulation that places it within the agent’s action space, sharing an ordinal scale with navigation, so that asking competes directly with acting at each decision point and help-seeking can be observed in intermediate states.
    • Two structurally distinct information-seeking modes emerge from the agent’s own ratings: mandatory (no viable branch) and opportunistic (residual uncertainty despite a leading candidate).
  • EN Key Points:
    • arXiv:2606.11349v1 Announce Type: new
    • Abstract: In hierarchical reasoning, failures often originate at intermediate decision points where the agent commits to a wrong branch without recognizing that…
    • Rather than treating clarification as an external uncertainty trigger, we propose ACTION-RATING, a formulation that places it inside the agent’s action space on…
    • Two structurally distinct information-seeking modes emerge from the agent’s own ratings: mandatory (no viable branch) and opportunistic (residual uncertainty de…

• Automated Mediator for Human Negotiation: Pre-Mediation via a Structured LLM Pipeline

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11379v1 Announcement Type: New.
    • Abstract: Pre-mediation, the preparatory phase preceding direct human negotiation, plays a critical role in achieving mutually beneficial agreements, yet is often omitted due to cost, time, and the limited availability of trained mediators.
    • We introduce an automated mediator for human negotiation, implemented as a structured pipeline of LLM modules, that supports pre-mediation in integrative negotiation settings.
    • The pipeline breaks down the preparation into dedicated modules for dialogue, preference prediction, response-level critique, and structured summarization, separating reasoning, generation, and evaluation to address the limitations of single-prompt approaches.
  • EN Key Points:
    • arXiv:2606.11379v1 Announce Type: new
    • Abstract: Pre-mediation, the preparatory phase preceding direct human negotiation, plays a critical role in achieving mutually beneficial agreements, yet is oft…
    • We introduce an automated mediator for human negotiation, implemented as a structured pipeline of LLM modules, that supports pre-mediation in integrative negoti…

• The pipeline decomposes preparation into specialized modules for dialogue, preference prediction, response-level critique, and structured summarization, separat…

• INFRAMIND: Infrastructure-Aware Multi-Agent Orchestration

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11440v1 Announcement Type: New.
    • Abstract: Existing multi-agent LLM orchestration methods, from brute-force integration to learned routers, select models and topologies based on task and model characteristics.
    • However, these methods do not consider the runtime state of the serving infrastructure.
    • On shared GPU clusters under concurrent load, this infrastructure blindness leads to systematic resource underutilization: preferred models accumulate deep request queues, while equally capable alternative models remain idle.
  • EN Highlights:
    • arXiv:2606.11440v1 Announce Type: new
    • Abstract: Existing multi-agent LLM orchestration methods, ranging from brute-force ensembles to learned routers, select models and topologies based on task and…
    • However, these methods do not consider the runtime state of the serving infrastructure
    • On shared GPU clusters under concurrent load, this infrastructure blindness causes systematic resource underutilization: preferred models accumulate deep reques…

• Forecasting Future Behavior as a Learning Task

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11445v1 Announcement Type: New.
    • Abstract: Trust in artificial intelligence systems often depends on an explanation of how they work, which is then used to predict their behavior on new inputs.
    • For large reasoning models (LRMs), this traditional route is particularly difficult to follow: explanation methods for single token generation do not naturally generalize to long trajectories, and the trajectories themselves are often not faithful when read as natural language.
    • We propose an alternative that bypasses the explanation step: treat behavior forecasting as a learnable task, and train behavior forecasters that operate on a single reasoning trajectory to make the same predictions that people usually seek from an explanation.
  • EN Highlights:
    • arXiv:2606.11445v1 Announce Type: new
    • Abstract: Trust in an AI system is often anchored by explanations of how it works, which one then uses to forecast its behavior on new inputs
    • For large reasoning models (LRMs), this conventional route is particularly difficult to follow: explanation methods for single token generations do not naturall…
    • We propose an alternative that bypasses the explanation step: treat behavior forecasting as a learnable task and train Behavior Forecasters that operates on a s…

• Search Discipline for Long-Horizon Research Agents

  • Publication Time: 2026-06-11 12:00 Beijing Time

• Abstract:- arXiv:2606.11522v1 Announce Type: new.

  • Abstract: Automated research agents now propose, evaluate, and select scientific candidates against a metric, and that metric is usually an aggregate reduced over a heterogeneous space of regions, slices, or groups.
  • We show that when scientific validity lives in that disaggregated structure, the aggregate can rank the wrong candidate first.
  • The headline number improves while the structure underneath inverts, so a decision made on the number accepts a candidate that quietly breaks the model.

• MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning

  • Published: 2026-06-11 12:00 Beijing Time
  • Abstract:- arXiv:2606.11537v1 Announce Type: new.
  • Abstract: Answering financial and tabular questions requires more than fluent reasoning: answers must be grounded in the exact facts, formulas, units, signs, and scales that support them.
  • A single misread cell or incorrect operation can silently produce a plausible but wrong result.
  • We introduce \textsc{MOCA-Agent}, a market-of-claims code agent that replaces free-form multi-agent debate with claim-level verification.

• SkillJuror: Measuring How Agent Skill Organization Changes Runtime Behavior

  • Published: 2026-06-11 12:00 Beijing Time
  • Abstract:- arXiv:2606.11543v1 Announce Type: new.
  • Abstract: Agent skills enhance Large Language Model (LLM) agents with procedural knowledge at inference time, but current benchmarks rarely distinguish between the content and the organization of skills.
  • We study this distinction through progressive disclosure, where a concise root file directs the agent to on-demand supporting resources, and compare it against a standardized, flat baseline.
  • We propose SkillJuror, a framework for evaluating skill-authoring paradigms through semantically controlled variants, matched multi-trial evaluation, and trajectory evidence, while keeping task knowledge fixed.

• Abstract: Agent Skills augment large language model (LLM) agents with procedural knowledge at inference time, but current benchmarks rarely distinguish what a S…

• We study this distinction through Progressive Disclosure, where a concise root file points agents to supporting resources on demand, and compare it with a norma…

• We present SkillJuror, a framework for evaluating Skill writing paradigms through semantically controlled variants, matched multi-trial evaluations, and traject…

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

• PoQ-Judge: A Multi-Architecture Evaluation Framework for Cost-Aware Proof-of-Quality in Decentralized LLM Inference

  • Published: 2026-06-11 12:00 Beijing Time
  • Abstract: [To be translated] - arXiv:2606.11196v1 Announce Type: new.
    • Abstract: Decentralized LLM inference networks need lightweight, reference-free quality evaluation for Proof of Quality (PoQ).
    • We present PoQ-Judge, a framework that trains dedicated judge models to score query-output pairs without ground-truth references.
    • We study three architectures across the quality-cost tradeoff: a TextCNN judge, a MiniLM cross-encoder, and a DeBERTa judge.
  • EN Highlights:
    • arXiv:2606.11196v1 Announce Type: new
    • Abstract: Decentralized LLM inference networks need lightweight, reference-free quality evaluation for Proof of Quality (PoQ)
    • We present PoQ-Judge, a framework that trains dedicated judge models to score query-output pairs without ground-truth references
    • We study three architectures across the quality-cost tradeoff: a TextCNN judge, a MiniLM cross-encoder, and a DeBERTa judge

• The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content

  • Published: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11198v1 Announce Type: new.
    • Abstract: Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM output, but the format of the injected content (distinct from its semantic relevance) can independently distort the model’s attention distribution.

• We identify and formalize a phenomenon we call the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3 times more attention per token than semantically equivalent natural language text ($\hat{o}$(KG) $\approx$ 0.70 vs.

  • $\hat{o}$(neutral) $\approx$ 0.25), compressing demonstration attention by up to 42%—regardless of whether the triples are relevant or noise.
  • EN Highlights:
    • arXiv:2606.11198v1 Announce Type: new
    • Abstract: Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content – distinct from its…
    • We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot…
    • $\hat{o}$(neutral) $\approx$ 0.25), compressing demonstration attention by up to 42% – regardless of whether the triples are relevant or noise

• NightFeats @ MMU-RAGent NeurIPS 2025: A Context-Optimized Multi-Agent RAG System for the Text-to-Text Track

  • Release Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11199v1 Announce Type: new.
    • Abstract: We present NightFeats, a structured multi-agent retrieval-augmented generation (RAG) system submitted to the MMU-RAGent competition at NeurIPS 2025, which won the best dynamic evaluation award for the text-to-text track.
    • This work, rather than aiming for benchmark maximization, proposes a principled pipeline that decomposes knowledge synthesis into three coordinated phases: retrieval, management, and composition, with each phase governed by explicit intermediate representations and handoff contracts.
    • Inspired by Agentic Context Engineering (ACE), the system introduces temporal-semantic reranking, bounded contradiction reconciliation, and citation-preserving composition as core architectural primitives.
  • EN Highlights:
    • arXiv:2606.11199v1 Announce Type: new
    • Abstract: We present NightFeats, a structured multi-agent retrieval-augmented generation (RAG) system submitted to the MMU-RAGent competition at NeurIPS 2025, w…
    • Rather than targeting benchmark maximization, this work proposes a principled pipeline that decomposes knowledge synthesis into three coordinated phases: retrie…
    • Inspired by Agentic Context Engineering (ACE), the system introduces temporal-semantic reranking, bounded contradiction reconciliation, and citation-preserving…

• Detecting AI-Generated Content on Social Media with Multi-modal Language Models

  • Release Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11200v1 Announce Type: new.
    • Abstract: Generative AI can create realistic images and videos that are increasingly disseminated on social media, often for spam, misinformation, manipulation, and fraud.

• Existing Artificial Intelligence Generated Content (AIGC) detection methods face challenges, including poor generalization to new-generation models, reliance on single modalities, and a lack of explainable interpretations.

• We propose a pipeline that mitigates these issues by continuously curating diverse multi-modal social media data and training a compact vision-language model for detection and explanation.

• EN Highlights:

  • arXiv:2606.11200v1 Announce Type: new
  • Abstract: Generative AI has enabled the creation of photorealistic images and videos that are increasingly disseminated on social media, often used for spam, mi…
  • Existing AI-generated content (AIGC) detection methods face challenges including poor generalization to new generation models, reliance on single modalities, an…
  • We present our pipeline that mitigates these issues by continuously curating diverse multi-modal social media data and training a compact vision-language model…

• One Jailbreak, Many Tongues: Learning Language-Insensitive Intention Representations for Multilingual Jailbreak Detection

  • Release Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11202v1 Announce Type: new.
    • Abstract: Large language models (LLMs) are increasingly deployed in applications for global multilingual users, yet safety training remains concentrated on dominant languages and has not developed in sync with multilingual capabilities, creating exploitable vulnerabilities for jailbreak attacks.
    • Current jailbreak defenses are primarily developed and evaluated in mainstream languages, and their effectiveness is limited by the lack of consistent multilingual supervision and representation dispersion caused by language variations.
    • To address this issue, we propose MLJailDe, a multilingual jailbreak detection framework designed to improve both multilingual robustness and cross-lingual generalization capabilities.
  • EN Highlights:
    • arXiv:2606.11202v1 Announce Type: new
    • Abstract: Large language models (LLMs) are increasingly deployed in applications for global multilingual users, yet safety training remains concentrated in domi…
    • Current jailbreak defenses are largely developed and evaluated in dominant languages, and their effectiveness is limited by the scarcity of aligned multilingual…
    • To address this issue, we propose MLJailDe, a multilingual jailbreak detection framework designed to improve both multilingual robustness and cross-lingual gene…

• LatticeBridge: Rare-Event Sequential Inference for Faithful Structured Sequence Synthesis

  • Release Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11203v1 Announce Type: new.
    • Abstract: Structured sequence generation often requires a model to satisfy multiple input-derived constraints in a single output.
    • Standard decoding methods may assign high probabilities to fluent continuations while placing low probability on continuations that jointly realize all desired anchors.
    • We study this mechanism as a rare-event sequential inference problem.
  • EN Highlights:

• arXiv:2606.11203v1 Announce Type: new

• Abstract: Structured sequence generation often requires a model to satisfy several input-derived constraints in a single output

• Standard decoding methods may assign high probability to fluent continuations while placing low mass on continuations that realize all required anchors jointly

• We study this regime as a rare-event sequential inference problem

• Benchmarking Large Language Models for Safety Data Extraction

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11204v1 Announce Type: new.
    • Abstract: Due to heterogeneous document formats and the limitations of traditional rule-based methods, accurately extracting structured information from Safety Data Sheets (SDS) remains challenging in the field of industrial safety.
    • This study benchmarks state-of-the-art Large Language Models (LLMs) for automated SDS data extraction, comparing text-based and multimodal processing pipelines.
    • We systematically evaluate four models: Gemini 1.5 Pro, GPT-4o, Claude 3.7 Sonnet, and Llama 3.1-70B, across three prompting strategies: zero-shot, few-shot, and chain-of-thought.
  • EN Highlights:
    • arXiv:2606.11204v1 Announce Type: new
    • Abstract: Accurate extraction of structured information from Safety Data Sheets (SDS) remains challenging in industrial safety due to heterogeneous document for…
    • This study benchmarks state-of-the-art Large Language Models (LLMs) for automated SDS data extraction, comparing text-based and multimodal processing pipelines
    • We systematically evaluate four models: Gemini 1.5 Pro, GPT-4o, Claude 3.7 Sonnet, and Llama 3.1-70B, across three prompting strategies: zero-shot, few-shot, an…

• Compatibility-Aware Dynamic Fine-Tuning for Large Language Models

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11206v1 Announce Type: new.
    • Abstract: Supervised Fine-Tuning (SFT) is the predominant paradigm for aligning Large Language Models (LLMs), yet it suffers from optimization instability and limited generalization.
    • Recent work attributes this problem to ill-conditioned gradient scaling and proposes Dynamic Fine-Tuning (DFT) to correct it at the token level.
    • However, DFT assumes all demonstrations are equally suitable learning targets, an assumption violated by the strong heterogeneity of large-scale instruction data, where mismatched demonstration strategies lead to high-variance updates at the sample level.
  • EN Highlights:
    • arXiv:2606.11206v1 Announce Type: new
    • Abstract: Supervised Fine-Tuning (SFT) is the predominant paradigm for aligning large language models (LLMs), yet it suffers from optimization instability and l…

• Recent work attributes this issue to pathological gradient scaling and proposes Dynamic Fine-Tuning (DFT) to correct it at the token level

• However, DFT assumes all demonstrations are equally suitable learning targets, an assumption violated by the strong heterogeneity of large-scale instruction dat…

• BioDivergence: A Benchmark and Evaluation Framework for Hidden Contextual Contradictions in Biomedical Abstracts

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11208v1 Announcement Type: new.
    • Abstract: Biomedical findings across different studies often seem to conflict, but many of these differences are context-dependent rather than true contradictions.
    • Variations in cohort, geography, assay protocol, disease subtype, and clinical setting can make both claims locally valid.
    • Existing NLI and scientific claim verification benchmarks reduce such cases to entailment, contradiction, or neutral, failing to capture the contextual structure behind the divergence.
  • EN Highlights:
    • arXiv:2606.11208v1 Announce Type: new
    • Abstract: Biomedical findings often seem to conflict across studies, but many of these differences are context-dependent rather than true contradictions
    • Variations in cohort, geography, assay protocol, disease subtype, and clinical setting can make both claims locally valid
    • Existing NLI and scientific claim-verification benchmarks reduce such cases to entailment, contradiction, or neutral, failing to capture the contextual structur…

• ProcessThinker: Enhancing Multi-modal Large Language Models Reasoning via Rollout-based Process Reward

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11209v1 Announcement Type: new.
    • Abstract: Visual question answering increasingly requires multi-step reasoning.
    • Recent post-training with reinforcement learning under verifiable rewards (RLVR) and Group Relative Policy Optimization (GRPO) can improve multimodal reasoning, but most methods rely on sparse, result-only rewards.
    • Therefore, they struggle to distinguish whether an incorrect answer results from a minor error late in the reasoning process or from an unhelpful trajectory from the very beginning.
  • EN Highlights:
    • arXiv:2606.11209v1 Announce Type: new
    • Abstract: Visual question answering increasingly requires multi-step reasoning
    • Recent post-training with reinforcement learning under verifiable rewards (RLVR) and Group Relative Policy Optimization (GRPO) can improve multimodal reasoning,…

• As a result, they struggle to tell whether an incorrect answer comes from a small mistake late in the reasoning or from an unhelpful trajectory from the start

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

• Restless bandits with imperfect binary feedback: PCL-indexability analysis and computation

  • Posted: 2026-06-11 12:00 Beijing Time
  • Abstract:
    • arXiv:2606.11192v1 Announcement Type: new.
    • We study restless bandits with binary latent states and imperfect binary feedback, motivated by opportunistic spectrum access with sensing errors.
    • For the associated belief-state model, we develop a partial conservation laws (PCL)-based analytical and computational framework for establishing indexability and evaluating the Whittle index, building upon verification theorems for restless bandits with discounted true states.
    • The framework analyzes the stochastic dynamics via an associated deterministic skeleton, renewal decompositions, and combinatorics on words.
  • EN Key Points:
    • arXiv:2606.11192v1 Announce Type: new
    • Abstract: We study restless bandits with binary latent states and imperfect binary feedback, motivated by opportunistic spectrum access with sensing errors
    • For the associated belief-state model, we develop a partial conservation laws (PCL)-based analytical and computational framework for establishing indexability a…
    • The framework analyzes the stochastic dynamics via an associated deterministic skeleton, renewal decompositions, and combinatorics on words

• To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blending

  • Posted: 2026-06-11 12:00 Beijing Time
  • Abstract:
    • arXiv:2606.11201v1 Announcement Type: new.
    • The wide deployment of LLMs has made model alignment necessary to make newly trained models safely and effectively respond to user instructions.
    • Among different methods, inference-time alignment is often cheaper as it intervenes (i.e., offers guidances) only during output generation.
    • Existing proposals apply guidances extracted from certain aligned models without properly assessing their reliability.
  • EN Key Points:
    • arXiv:2606.11201v1 Announce Type: new
    • Abstract: The wide deployment of LLMs has made model alignment necessary to make newly trained models safely and effectively respond to user instructions
    • Among different methods, inference-time alignment is often cheaper as it intervenes (i.e., offers guidances) only during output generation
    • Existing proposals apply guidances extracted from certain aligned models without properly assessing their reliability

• Dual-Stance Evaluation of Sycophancy: The Structure of Agreement and the Limits of Intervention

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11205v1 Announcement Type: New.
    • Abstract: Activation steering can alter LLM behavior, but standard evaluations do not typically test whether a direction that reduces sycophancy also suppresses agreement with factually correct statements.
    • We introduce dual-stance evaluation, which tests both stances for each topic, and apply it to centroid-difference steering on Llama-3-8B-Instruct.
    • We find a dissociation: the model represents sycophantic and factual agreement in geometrically distinct subspaces, yet the steering direction projects equally onto both and cannot target either one discriminately.
  • EN Key Points:
    • arXiv:2606.11205v1 Announce Type: new
    • Abstract: Activation steering can shift LLM behaviour, but standard evaluations do not typically test whether a sycophancy-reduction direction also suppresses a…
    • We introduce dual-stance evaluation, which tests both stances of each topic, and apply it to centroid-difference steering on Llama-3-8B-Instruct
    • We find a dissociation: the model represents sycophantic and factual agreement in geometrically distinct subspaces, yet the steering direction projects equally…

• Few-Shot Resampling for Scalable Statistically-Sound Data Mining

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11235v1 Announcement Type: New.
    • Abstract: A key step in knowledge discovery is the evaluation of data mining results.
    • In various applications, including pattern mining, graph analysis, and others, this step includes evaluating the statistical significance of the results to avoid spurious findings due solely to noise or random fluctuations in the data.
    • While specialized procedures have been developed for some specific applications, resampling-based approaches are widely used, particularly for complex analyses where analytical results cannot be derived.
  • EN Key Points:
    • arXiv:2606.11235v1 Announce Type: new
    • Abstract: A key step in knowledge discovery is the evaluation of data mining results
    • In several applications, including pattern mining, graph analysis, and others, this step includes the evaluation of the statistical significance of the results,…
    • While specialized procedures have been developed for some specific applications, resampling-based approaches are widely used, in particular for complex analyses…

• ProHiFlo: Hierarchical Flow Matching with Functional Guidance for De Novo Protein Generation

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11243v1 Announcement Type: New.
    • Abstract: De novo protein generation holds transformative potential for therapeutic design, enzyme engineering, and synthetic biology.

• While diffusion-based and flow matching approaches have achieved progress, they typically operate at a single resolution and lack mechanisms for incorporating functional constraints.

• We introduce ProHiFlo, a hierarchical flow matching framework with three innovations: (1) coarse-to-fine generation that models backbone geometry before refining to full-atom coordinates, thus reducing computational cost while maintaining accuracy; (2) functional guidance that leverages pretrained predictors to steer a generation toward desired properties without retraining; and (3) an adaptive SE(3)-equivariant architecture for efficient multi-scale processing.

  • EN Highlights:
    • arXiv:2606.11243v1 Announce Type: new
    • Abstract: De novo protein generation has transformative potential in therapeutic design, enzyme engineering, and synthetic biology
    • While diffusion-based and flow matching approaches have achieved progress, they typically operate at single resolution and lack mechanisms for incorporating fun…
    • We introduce ProHiFlo, a hierarchical flow matching framework with three innovations: (1) coarse-to-fine generation that models backbone geometry before refinin…

• Physics-informed generative AI for semiconductor manufacturing: Enforcing hard physical constraints in generative models by construction

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11247v1 Announcement Type: new.
    • Abstract: Generative models are increasingly used to propose designs, data, and control actions for physical systems, yet many such systems are governed by hard physical constraints rather than perceptual plausibility.
    • Semiconductor manufacturing provides a demanding test case: generated masks, layouts, synthetic defect data, and process recipes must obey lithography, transport, reaction, and device physics constraints, as physically invalid samples are not just low quality but unusable.
    • This Perspective argues that semiconductor manufacturing exposes a broader computational-science challenge, namely that generative AI for constrained physical domains must be physics-informed by construction, and not just corrected by post-hoc filtering.
  • EN Highlights:
    • arXiv:2606.11247v1 Announce Type: new
    • Abstract: Generative models are increasingly used to propose designs, data, and control actions for physical systems, yet many such systems are governed by hard…
    • Semiconductor manufacturing provides a demanding test case: generated masks, layouts, synthetic defect data, and process recipes must obey lithography, transpor…
    • This Perspective argues that semiconductor manufacturing exposes a broader computational-science challenge, namely that generative AI for constrained physical d…

• Mechanical Field Networks: Structured Neural Dynamics for Multivariate Systems

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11251v1 Announcement Type: new.
    • Abstract: Many multivariate dynamical systems are observed only through their trajectories, which hides the mechanisms that govern their joint dynamics.

• Existing approaches can impose interpretable dynamics or learn flexible state transitions, but the resulting interaction structure is typically either specified in advance or implicit in the learned dynamics.

• We introduce MF-Net, a recurrent dynamical model that represents all variables in a shared field state and updates this state through a learned relational law.

• EN Highlights:

  • arXiv:2606.11251v1 Announce Type: new
  • Abstract: Many multivariate dynamical systems are observed only through trajectories, leaving the mechanisms governing their joint dynamics hidden
  • Existing approaches can impose interpretable dynamics or learn flexible state transitions, yet the resulting interaction structure is typically either specified…
  • We introduce MF-Net, a recurrent dynamical model that represents all variables in a shared field state and updates this state through a learned relation law

• Bernstein-Schur Kernels: Random Features by Sketched Modulation and Radial Randomization

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11255v1 Announcement Type: New.
  • Abstract: Bernstein-Schur kernels are products of a finite-feature kernel (one with an explicit finite-dimensional feature map) and a completely monotone translation-invariant kernel. These non-stationary kernels, which are intermediate between shift-invariant and dot-product templates, typically leverage random features. However, in general, neither Bochner sampling nor polynomial sketching can be directly applied to the full kernel.
  • We provide a random feature construction for the entire class that \emph{randomizes both factors}: it sketches the finite modulation and randomizes the completely monotone radial factor by sampling the latter’s one-dimensional Bernstein-Widder measure, and then applies Gaussian random Fourier features (whose frequencies remain $d$-dimensional).
  • The feature dimension is $Dm$, set by the sketch size $m$ and the radial draw count $D$, and is unaffected by the $O(d^2)$ size of the exact modulation features.
  • EN Highlights:
    • arXiv:2606.11255v1 Announce Type: new
    • Abstract: Bernstein–Schur kernels are products of a finite-feature kernel (one with an explicit finite-dimensional feature map) and a completely monotone shift…
    • We give one random-feature construction for the whole class that \emph{randomizes both factors: it sketches the finite modulation and randomizes the completely…
    • The feature dimension is then $Dm$, set by the sketch size $m$ and the radial-draw count $D$, free of the $O(d^2)$ size of the exact modulation feature

• Loss Landscape Diagnosis for Gradient-Based Gray-Scott System Inversion: Disentangling the Roles of PINN Components

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11258v1 Announcement Type: New.
  • Abstract: Gradient-based inversion of reaction-diffusion systems is typically achieved through surrogate models or Physics-Informed Neural Networks (PINNs), while the most direct route—backpropagation through the PDE structure itself—has been largely avoided.

• We use this direct route as a diagnostic probe, backpropagating a steady-state loss through unrolled Gray-Scott simulation to recover its parameters, without surrogate or neural network augmentation.

• Optimization fails to converge, and plotting the landscape directly locates the failure in its geometry—flat plateaus with no gradient signal, bounded by steep cliffs that align with bifurcation boundaries. This structure recurs in the loss function and is inherited, but the gradient is routed to the parameters.

  • EN Highlights:
    • arXiv:2606.11258v1 Announce Type: new
    • Abstract: Gradient-based inversion of reaction-diffusion systems is typically approached via surrogate models or physics-informed neural networks (PINNs), while…
    • We pursue this direct route as a diagnostic probe, backpropagating a steady-state loss through unrolled Gray-Scott simulation to recover its parameters, with no…
    • Optimization fails to converge, and plotting the landscape directly locates the failure in its geometry – flat plateaus with no gradient signal, bounded by sha…

• PermDoRA – Understanding Adapter Interference in Language Models: Limits of Parameter-Space Geometry

  • Publication Time: 2026-06-11 12:00 Beijing Time
  • Abstract: - arXiv:2606.11262v1 Announcement Type: New.
    • Abstract: Access control in large language models (LLMs) requires modular mechanisms to enable domain-specific behavior without retraining or cross-domain interference.
    • A common hypothesis is that interference during adapter composition arises from the overlap of linear parameter updates, suggesting that enforcing orthogonality or directional independence should improve multi-domain performance.
    • We test this hypothesis using DoRA-RBAC, a hierarchical adapter composition framework based on weight-decomposed low-rank adaptation.
  • EN Highlights:
    • arXiv:2606.11262v1 Announce Type: new
    • Abstract: Access control in large language models (LLMs) requires modular mechanisms to enable domain-specific behavior without retraining or cross-domain inter…
    • A common hypothesis is that interference during adapter composition arises from overlap in linear parameter updates, suggesting that enforcing orthogonality or…
    • We test this hypothesis using DoRA-RBAC, a hierarchical adapter composition framework based on weight-decomposed low-rank adaptation