System translated (Gemini)

🤖 AI 速览

OpenAI has released GPT-6 Sol/Luna, cutting API prices by 50% and unveiling prompt caching diagnostics with breakpoint controls. A case study from Parallel demonstrates that Astra can reduce research time and costs by half. Multiple papers on inference optimization are focusing on sparse attention …
📋 文章元数据
发布时间
2026-09-23
类型
ai-daily
字数
7199
阅读时长
34 min

2026-09-23 AI Daily | GPT-6 Dual Models Rewrite Inference Costs: Cache Diagnostics, Breakpoint Control & Half-Price API Link to heading

OpenAI releases GPT-6 Sol/Luna, cutting API prices by 50% and detailing prompt cache diagnostics and breakpoint control. A Parallel case study shows Astra can halve research time and costs. Multiple papers on inference optimization focus on sparse attention for long contexts and MoE routing, shifting the service competition to cost engineering. Amazon’s move to block Muse highlights the distribution boundary dispute between platforms and AI assistants.

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

Today’s Watch List features two main threads worth a deep dive. The first is the engineering progress of GPT-6: OpenAI’s blog has seen a flurry of updates on the positioning of the Sol/Luna dual models, prompt cache diagnostics, breakpoint control, and a case study from Parallel on using Astra to cut research time and costs in half. Teams focused on inference infrastructure should pay close attention to cache hit rates and cost structures, which represent the real-world efficiency calculus beyond model capabilities. The second is a batch of inference optimization papers emerging on arXiv: RBS-Attention tackles the failure of sparse attention in long-context prefilling, and Attention-Aware Routing enhances MoE routing with attention features. These are an excellent technology radar for teams optimizing large model service performance. On the business side, the discussions around Amazon blocking Muse and the Jevons Paradox are also worth reading to understand platform moats and the boundary conditions for compute demand.

🌐 AI Hot Topics on X Link to heading

Topic 1: CopilotKit Launches OpenMuse Open-Source Personal AI Assistant Link to heading

  • Category: AI · News
  • Overview: Trending for: 3 hours ago, Related posts: 381
  • What it is: CopilotKit has launched OpenMuse, an open-source personal AI assistant.
  • Why it matters: This move introduces the open-source model to the personal AI assistant space, potentially fostering the development of more transparent, customizable, and privacy-focused assistants, and intensifying competition in the field.
  • Discussion summary: Discussions on X are focused on the capability gap between OpenMuse and existing closed-source assistants, the significance of open source for privacy and customization, and its practical usability and ecosystem prospects.

Topic 2: Unreal Agent Open-Source Harness Cuts AI Coding Costs by 39% Link to heading

  • Category: AI · Other
  • Overview: Trending for: , Related posts: 149
  • What it is: The open-source project Unreal Agent has released a testing harness that reportedly reduces AI coding costs by 39%.
  • Why it matters: The framework directly targets the high cost of AI programming agents. If this cost reduction is reproducible, it will improve the economic viability of AI coding and accelerate its adoption in development workflows.
  • Discussion summary: The focus on X is on the credibility of the 39% cost reduction claim, its applicability across different models and tasks, and the potential impact of an open-source solution on the pricing models of existing commercial AI coding tools.

Topic 3: Meta’s Muse AI Agent Tops App Charts and Boosts Stock Link to heading

  • Category: AI · News
  • Overview: Trending for: 8 hours ago, Related posts: 4,700
  • What it is: Meta’s Muse AI agent has topped the app store charts, driving a rise in Meta’s stock price.
  • Why it matters: This indicates that AI agents are becoming a key competitive product for tech giants in the consumer space, with their market performance now directly impacting company valuations and confidence in AI strategy.
  • Discussion summary: The main topics on X are whether its chart-topping performance is driven by genuine user demand or marketing, whether it can sustain its lead over competitors like ChatGPT and Gemini, and Meta’s long-term advantages and risks in AI applications and privacy.

Topic 4: White Circle Open-Sources Halo for Faster Model Training on Hugging Face Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 1,000
  • What it is: White Circle has open-sourced the Halo framework on Hugging Face, which reportedly accelerates distributed large language model training by up to 2.8 times.
  • Why it matters: This framework directly addresses the core bottlenecks of cost and time in large model training. If the speedup is confirmed, it would significantly improve computational efficiency and lower the barrier to entry for training.
  • Discussion summary: Discussions on X are focused on the real-world reproducibility of the 2.8x speedup, its advantages compared to existing frameworks like DeepSpeed, and its applicability across different models and hardware configurations.

Topic 5: AI Guide Makes Product Launch Videos Easy for Beginners Link to heading

  • Category: AI · News
  • Overview: Trending for: 16 hours ago, Related posts: 4,200
  • What it is: An AI video tool demonstrated how to clone a founder’s likeness and voice from a single photo and a short voice clip, automatically generating a product launch video complete with subtitles and a timeline.
  • Why it matters: This indicates that AI video generation, voice cloning, and automated post-production capabilities have reached a practical stage, significantly lowering the barrier and cost of producing product launch videos.
  • Discussion overview: On X, there is both amazement at the efficiency of “zero-barrier video creation” and skepticism about the actual value and homogenization of the vast number of AI tools. Some posts highlight its marketing effectiveness, while others urge a focus on the tangible results the tool produces.

Topic 6: OpenAI Launches Faster, Cheaper GPT-6 Sol and Luna Models Link to heading

  • Category: AI · News
  • Overview: Trending since: 5 hours ago, Related posts: 34,000
  • What it is: OpenAI released the GPT-6 Sol and GPT-6 Luna models, with API prices 50% lower than GPT-5.6, and they are now available on ChatGPT Work and Codex.
  • Why it matters: This marks a shift in the competitive focus from pure model capabilities to task-level pricing and inference costs. It could significantly lower the adoption barrier for AI agents and continuous inference, intensifying the cost-performance competition with Anthropic and open-source models.
  • Discussion overview: Opinions on X are sharply divided: one side sees the price reset and decreased inference costs as a major turning point, while the other questions the rapid frequency of model releases and the limited improvements in user experience, sparking discussions about an AI bubble and subscription fatigue. At the same time, users are comparing these releases to competing models launched the same day, such as Claude Opus 5.5, the open-source MiMo-V2.6-Pro, and Grok 4.7.

Topic 7: Pexo’s Mark to Fix Tool Transforms AI Video Creation Link to heading

  • Category: AI · Other
  • Overview: Trending since:, Related posts: 235
  • What it is: Pexo launched its Mark to Fix tool, which allows users to mark problematic areas in AI-generated videos and apply targeted fixes.
  • Why it matters: This tool enhances the fine-grained control over AI videos, reduces post-generation manual editing costs, and provides a practical boost to the AI video creation workflow.
  • Discussion overview: Key discussion points include the accuracy of its local repair feature, its impact on professional video workflows, and whether it offers a distinct advantage compared to other AI video editing tools.

Topic 8: Meta AI Chief Jokes About Muse Logo Dodging Butthole Comparisons Link to heading

  • Category: AI · Other
  • Overview: Trending since: 17 hours ago, Related posts: 2,600
  • What it is: The head of Meta AI joked that the Muse logo dodged being compared to the shape of a butt/anus.
  • Why it matters: The topic shows that the visual branding of AI products can become a focal point of public discussion, and the executive’s humorous response was seen as a way to defuse the sensitivity of the design controversy.
  • Discussion overview: The main focus on X is on whether AI logos are prone to vulgar interpretations and how brand design should balance recognizability with avoiding unfortunate associations. Some users feel that this type of discussion distracts from the technology itself.

Today’s AI Public Opinion Summary on X Link to heading

The main theme of today’s AI discussions on X is the concurrent rise of open-source, cost-reducing tools and consumer-grade AI applications. OpenMuse, Unreal Agent, and Halo claim to bring greater transparency or significant efficiency improvements in the domains of personal assistants, coding, and training, respectively. Meanwhile, Meta Muse’s rise to the top and OpenAI’s price cuts indicate that the competition among tech giants is shifting toward application rankings and task-level costs. The public consensus welcomes AI with lower barriers to entry, greater customizability, and more control, agreeing that AI agents and video generation have reached a practical stage. The main points of contention are whether metrics like “39% cost reduction” and “2.8x speed-up” are reproducible, whether Meta’s chart-topping performance is due to genuine demand or a marketing push, and whether the rapid pace of releases actually improves user experience. Potential risks include the possibility of open-source and commercial projects exaggerating their effectiveness and eroding trust. AI likeness and voice cloning introduce deepfake and privacy risks. Brand design controversies could also distract from technological discussions, while rapid price reductions and frequent releases heighten fears of an AI bubble and subscription fatigue.

💡 Influencer Insights Link to heading

No influencer insights available for today. We recommend reading the in-depth content from the Watch List.

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

Timeframe: Last 3 days; covers 22 sources; 37 total updates

All-In Podcast (A_full) Link to heading

  • Blake Scholl: Why Plane Speed Stalled, Supersonic Commercial Flight, & Revolutionizing the Engine
    • Published: 2026-09-22 17:00 (Beijing Time)
    • Abstract: - IREN is a vertically integrated AI cloud platform that provides data centers, computing, and software for AI training and inference.
  • Oracle connects data, applications, and infrastructure to turn AI into business outcomes and deliver optimized flexibility, choice, and control as AI evolves.
  • EY helps tech innovators from startup to exit to large enterprise.
  • We are committed to giving everyone the tools to reach their full potential and ensuring that all people can benefit from technology.
  • Keel Infrastructure has the power, land, and connectivity for high-performance computing and AI to run, backed by secure energy assets and grid interconnections across North America.
    • EN Highlights:
      • (0:00) Welcome Blake Scholl
      • (5:11) Solving the sonic boom & legalizing supersonic flight
      • (7:59) The pivot to powering data centers
      • (10:38) Jason joins: data center demand, private jets & what winning looks like

Stratechery by Ben Thompson (A_full) Link to heading

  • Amazon Blocks Muse, Amazon’s Moat, Aggregator v Aggregator
    • Publication Time: 2026-09-22 18:00 Beijing Time
    • Summary: - Amazon predictably blocked Muse, but there is room for a deal based on the reality that Amazon’s physical world investments are an AI moat.
      • $15/month or $150/year.
      • Substantive analysis of the day’s news via three weekly emails or podcasts.
      • Strategy Interviews.
      • Interviews with leading public company CEOs, private company founders, and discussions with fellow analysts.
    • EN Highlights:
      • Amazon predictably blocked Muse, but there is room for a deal based on the reality that Amazon’s physical world investments are an AI moat.

OpenAI Blog (A_full) Link to heading

  • Better prompt caching for GPT-6

    • Publication Time: 2026-09-23 05:00 Beijing Time
    • Summary: - Learn how GPT-6 improves prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and controls that reduce latency and costs.
      • This article from the OpenAI Blog explains how GPT-6’s better prompt caching is shaping the broader AI and infrastructure landscape.
      • Following GPT-6’s better prompt caching, it also has practical implications for founders, operators, and investors.
    • EN Highlights:
      • Learn how GPT-6 improves prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and controls that reduce latency and costs.

  • Introducing GPT-6 Sol and Luna

    • Publication Time: 2026-09-23 02:00 Beijing Time
    • Summary: - Meet GPT-6 Sol and Luna, two models that bring frontier intelligence to everyday work with different balances of capability and cost.
      • This article from the OpenAI Blog explains how the launch of GPT-6 Sol and Luna is shaping the broader AI and infrastructure landscape.
      • Following the launch of GPT-6 Sol and Luna, it also has practical implications for founders, operators, and investors.
    • EN Highlights:
      • Meet GPT-6 Sol and Luna, two models that bring frontier intelligence to everyday work with different balances of capability and cost.

  • Parallel cut research time and cost in half with GPT‑6 Astra

    • Publication Time: 2026-09-22 20:00 Beijing Time

  • Summary: - GPT-6 Astra allowed Parallel’s agents to research and synthesize labor-market data in half the time and at half the cost.

    • Compared to previous models, GPT-6 Astra allowed Parallel’s agents to research and synthesize labor market data in half the time and at half the cost.
    • Using GPT-6 Astra cut research time and cost for Parallel in half.

  • EN Highlights:

    • GPT-6 Astra allowed Parallel’s agents to research and synthesize labor-market data in half the time and at half the cost vs
    • prior models.

  • Priorities and principles for effective third party assessments

    • Published: 2026-09-22 08:00 Beijing Time
    • Summary: - OpenAI outlines priorities and principles for rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards.
      • This article from the OpenAI blog explains how the priorities and principles for effective third-party assessments shape the broader AI and infrastructure landscape.
      • It also provides founders, operators, and investors with the practical implications of following the priorities and principles for effective third-party assessments.
    • EN Highlights:
      • OpenAI outlines priorities and principles for rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards.

Two Minute Papers (B_intro+search) Link to heading

  • Yes, Jev Is Insane, But There’s A Catch
    • Published: 2026-09-22 17:32 Beijing Time
    • Summary: - ❤️ Check out Lambda here and sign up for their GPU Cloud:.
      • Previous papers, implementations, models:.
      • Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi.
      • Yes, Jev is insane, but there’s a catch.
    • EN Highlights:
      • ❤️ Check out Lambda here and sign up for their GPU Cloud:
      • Previous papers, implementations, models:
      • 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
      • Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stankye, Shawn Becker, Steef…

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

  • RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models

    • Published: 2026-09-22 12:00 Beijing Time
    • Summary: - arXiv:2609.20971v1 Announcement type: new.
      • Abstract: Long-context large language model inference is increasingly constrained by prefill, where dense self-attention processes the entire prompt before generation begins.
      • Sparse block selection can reduce this cost, but block centroids can hide highly relevant tokens among many irrelevant ones.
      • We call this failure mode “mean dilution” and propose RBS-Attention, a training-free sparse prefill method with two complementary selection branches.

  • EN Highlights:

    • arXiv:2609.20971v1 Announce Type: new
    • Abstract: Long-context large language model inference is increasingly limited by prefill, where dense self-attention processes the entire prompt before generati…
    • Sparse block selection can reduce this cost, but a block centroid may hide a highly relevant token among many irrelevant ones
    • We call this failure mode mean dilution and propose RBS-Attention, a training-free sparse-prefill method with two complementary selection branches

  • Attention-Aware Routing: Coupling Routing and Attention in MoEs

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.20974v1 Announcement Type: New.
      • Abstract: In Mixture-of-Experts language models, the router typically utilizes limited contextual information to select and weight experts based on the token’s hidden state.
      • We propose Attention-Aware Routing (AAR), which augments the router with temporal and spectral features extracted from a sliding window of attention weights that represent a summary of the model’s contextual state, separate from the hidden state.
      • Keeping the base transformer entirely frozen, we train only the routing parameters, isolating routing as the sole variable.
    • EN Highlights:
      • arXiv:2609.20974v1 Announce Type: new
      • Abstract: In Mixture-of-Experts language models, the router typically selects and weights experts based on the token’s hidden state, utilizing limited contextua…
      • We propose Attention-Aware Routing (AAR), which augments the router with temporal and spectral features extracted from a sliding window of attention weights tha…
      • Keeping the base transformer entirely frozen, we train only the routing parameters, isolating routing as the sole variable

  • CaLR: Causal Latent Revision for Robust Diffusion Reasoning

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.20981v1 Announcement Type: New.
      • Abstract: Autoregressive (AR) models suffer from the local greediness problem, while diffusion language models (DLMs) often lack the strict causal structure required for reasoning.
      • To combine the advantages of both and overcome their shortcomings, we propose Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrained latent optimization.
      • By employing a causal topological matrix (CTM) from an expert model and implicit differentiation, CaLR performs gradient-guided “thought revision” to enforce logical consistency, thereby enabling dynamic self-correction of intermediate steps during the parallel generation process.
    • EN Highlights:
      • arXiv:2609.20981v1 Announce Type: new
      • Abstract: Autoregressive (AR) models suffer from local greediness, while diffusion language models (DLMs) often lack the strict causal structure required for re…

  • To combine the advantages and overcome the drawbacks of the dual, we propose Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrain…

  • By adopting a causal topology matrix (CTM) from an expert model and implicit differentiation, CaLR performs gradient-guided ``thought revision" to enforce logic…

  • LoRA Enhanced Contrastive Learning with SAS Vision Transformers

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.21061v1 Announce Type: new.
      • Abstract: Automatic target recognition (ATR) with synthetic aperture sonar (SAS) supports advanced naval capabilities, but deep learning is constrained by the scarcity of target images, background clutter, and human-in-the-loop evaluation.
      • We adapt DINOv3 Vision Transformer (ViT) models to underwater SAS ATR using a three-stage parameter-efficient framework.
      • Stage 1 uses Low-Rank Adaptation (LoRA) while freezing the ViT backbone, bridging the gap between natural-image pretraining and underwater acoustic propagation.
    • EN Highlights:
      • arXiv:2609.21061v1 Announce Type: new
      • Abstract: Automatic target recognition (ATR) with synthetic aperture sonar (SAS) supports advanced naval capabilities, but deep learning is constrained by scarc…
      • We adapt DINOv3 Vision Transformer (ViT) models to underwater SAS ATR using a three-stage parameter-efficient framework
      • Stage 1 uses Low-Rank Adaptation (LoRA) while freezing the ViT backbone, bridging the gap between natural-image pretraining and underwater acoustic propagation

  • Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.21096v1 Announce Type: new.
      • Abstract: In this work, we examine the topology of information flow patterns within attention graphs to effectively distinguish between hallucinated and non-hallucinated responses.
      • We analyze the Forman-Ricci curvature to identify structural patterns that indicate information bottlenecks in attention graphs.
      • We then introduce a method to capture semi-local and global information flow features of attention heads associated with hallucinatory responses.
    • EN Highlights:
      • arXiv:2609.21096v1 Announce Type: new
      • Abstract: In this work, we examine the topology of information flow patterns within attention graphs to effectively distinguish hallucinated from non-hallucinat…
      • We analyze the Forman-Ricci curvature to identify structural patterns indicating information bottlenecks in attention graphs

  • We then introduce a method that captures both semi-local and global information-flow characteristics of attention heads associated with hallucinated responses

  • Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models

    • Release Date: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.21113v1 Announce Type: new.
      • Abstract: Fine-tuning has emerged as a widely adopted approach for adapting LLMs to a variety of downstream tasks.
      • However, how it reshapes their internal mechanisms remains poorly understood.
      • To address this, we investigate how fine-tuning alters internal representations in LLMs, including attention patterns and layer-wise activations, and examine whether these changes are related to EAP-identified task-relevant components (e.g., attention heads and logic-level activations), thereby driving task performance.
    • EN Key Points:
      • arXiv:2609.21113v1 Announce Type: new
      • Abstract: Fine-tuning has emerged as a widely adopted approach for adapting LLMs to a variety of downstream tasks
      • However, how it reshapes their internal mechanisms remains poorly understood
      • To address this, we investigate how fine-tuning alters internal representations in LLMs, including attention patterns and layer-wise activations, and examine wh…

  • TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers

    • Release Date: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.21139v1 Announce Type: new.
      • Abstract: Replacing attention in a pretrained language model is a compatibility problem: a plausible substitute may alter representations expected by later layers.
      • TinyCeNN-LM introduces a \emph{quality-gated post-training conversion} framework using CeNN-inspired cellular-recurrent layers with bounded local processing, compact recurrent memory, routing, fusion, and acceptance or rollback validation.
      • Three implementations are studied: Integrated Memory, MemoryFusion, and PDelta3-GDN2-CLVR+Local32.
    • EN Key Points:
      • arXiv:2609.21139v1 Announce Type: new
      • Abstract: Replacing attention in a pretrained language model is a compatibility problem: a plausible substitute may alter representations expected by later laye…
      • TinyCeNN-LM introduces a \emph{quality-gated post-training conversion} framework using CeNN-inspired cellular-recurrent layers with bounded local processing, co…
      • Three implementations are studied: Integrated Memory, MemoryFusion, and PDelta3-GDN2-CLVR+Local32

  • Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake

  • Publication Time: 2026-09-22 12:00 Beijing Time

    • Abstract: - arXiv:2609.21149v1 Announcement Type: new.
      • Abstract: Before patients can use AI-assisted psychiatric intake systems, health systems need practical methods to regularly evaluate these tools against their quality-assured clinical standards.
      • Because clinicians may use different intake styles, the evaluation for this task must (1) support comparison across different interview methods, (2) minimize the burden on clinicians, and (3) measure clinically relevant performance for the health systems deploying these technologies.
      • We propose a clinician-grounded evaluation platform built around a memory-augmented patient simulator for open-ended AI interviewing, InterviewPlayground.
    • EN Highlights:
      • arXiv:2609.21149v1 Announce Type: new
      • Abstract: Before patients can use AI-assisted psychiatric intake systems, health systems need practical ways to routinely evaluate these tools against their cli…
      • Because clinicians may use different intake styles, evaluation for this task must (1) support comparison across interviewing approaches, (2) minimize clinician…
      • We present a clinician-grounded evaluation platform built around a memory-augmented patient simulator for open-ended AI interviewing, InterviewPlayground

  • Can Agents Design Better Chips with a Higher Level Abstraction?

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.21157v1 Announcement Type: new.
      • Abstract: Large Language Model (LLM) agents are increasingly being explored for chip design, but most existing methods operate directly on RTL.
      • We ask whether agents can design better chips by leveraging higher-level abstractions.
      • We compare direct RTL design, agent-based HLS design, post-compiler HLS refinement, and post-HLS RTL refinement, and combine agent-based HLS design with post-HLS RTL refinement as Agent-based HLS with RTL Refinement (AHRR).
    • EN Highlights:
      • arXiv:2609.21157v1 Announce Type: new
      • Abstract: Large Language Model (LLM) agents are increasingly being explored for chip design, but most existing approaches operate directly at RTL
      • We ask whether agents can design better chips by leveraging higher-level abstractions
      • We compare Direct RTL Design, Agent-based HLS Design, Post-Compiler HLS Refinement, and Post-HLS RTL Refinement, and combine Agent-based HLS Design with Post-HL…

  • SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.21165v1 Announcement Type: new.
      • Abstract: Off-target protein binding is a major source of adverse reactions for small molecule drugs, but most structure-based molecular design methods focus on de novo generation of selective compounds rather than improving the selectivity of existing, well-characterized drugs.

  • We introduce specificity optimization (SpecOpt), a molecular design task that seeks to make limited structural modifications to existing compounds to increase their binding preference for the intended target over known off-targets, while preserving their structural and drug-like properties.

  • To enable systematic evaluation, we constructed a ChEMBL-derived benchmark based on compound-target interaction data, identifying intended targets through curated drug mechanism-of-action annotations and off-targets through measured activity.

    • EN Key Points:
      • arXiv:2609.21165v1 Announce Type: new
      • Abstract: Off-target protein binding is a major source of adverse effects for small-molecule drugs, yet most structure-based molecular design methods focus on g…
      • We introduce specificity optimization (SpecOpt), a molecular design task that seeks constrained structural modifications to an existing compound that increase i…
      • To enable systematic evaluation, we construct a ChEMBL-derived benchmark from compound-target interaction data, identifying intended targets through curated dru…

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

  • Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22090v1 Announce Type: new.
      • Abstract: An LLM that produces response patterns associated with human psychological effects is different from an LLM that possesses such biases.
      • We present PsyAgentBench, a benchmark that re-runs classic psychology experiments on LLM agents, constructed to separate these experiments: each paradigm is run either with the paradigm explicitly labeled in the prompt (named) or as a regular task (blind), and on either a literal textbook version of the task (canonical) or a structurally matched variant to reduce lexical and scenario overlap with possible training data (counterfactual), crossed with role-playing manipulations.
      • Across five completed paradigms, evaluated on up to three open-weight model families with 41,904 trials released, apparently human-like effects arise through qualitatively different pathways rather than a single sensitivity: gating by explicitly-covered paradigm labels (Asch conformity, 0% blind to 83.3% named on gpt-oss-120B), knowledge-dependent signal-dependence (anchoring, with near-zero precision on fact-based vs. near-total on invented quantities, a pattern equally consistent with rational use of the only available signal), amplification of new content under a label (framing), robust absence (sunk cost), and safety-mediated choices where refusal is itself the main finding (minimal group assignment).
    • EN Key Points:
      • arXiv:2609.22090v1 Announce Type: new
      • Abstract: An LLM producing the response pattern associated with a human psychological effect is not the same claim as the LLM possessing that bias
      • We present PsyAgentBench, a benchmark that re-runs classic psychology experiments on LLM agents under a factorial design built to separate these: each paradigm…
      • Across five completed paradigms, evaluated on up to three open-weight model families with 41,904 trials released, apparently human-like effects arise through qu…

  • Memory That Looks Forward: A Zero-Inference Prospective Term for Personal Memory Retrieval

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Summary:
      • arXiv:2609.22091v1 Announcement Type: new.
      • Abstract: Retrieval over a personal memory store is retrospective: it surfaces content resembling the query and is blind to actions the user has committed to perform.
      • We describe a prospective term for memory retrieval that requires no inference at query time.
      • Commitments are held in an explicit ledger as dated or trigger-conditioned entries; memory items linked to a triggered entry receive a salience boost and are multiplicatively blended into embedding-based retrieval so that relevance remains sovereign.
    • EN Key Points:
      • arXiv:2609.22091v1 Announce Type: new
      • Abstract: Retrieval over a personal memory store is retrospective: it surfaces what resembles the query, and it is blind to what the user has committed to do
      • We describe a prospective term for memory retrieval that costs no inference at query time
      • Commitments are held in an explicit ledger as dated or trigger-conditioned entries; memory items linked to a firing entry receive a salience boost, blended mult…

  • Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Summary:
      • arXiv:2609.22094v1 Announcement Type: new.
      • Abstract: Content moderation systems traditionally entangle multimodal understanding with policy-specific classification, requiring full pipeline retraining for every policy change and suffering from label scarcity due to the inability to meaningfully augment multimedia.
      • We propose Summarize-Judge-Refine (SJR), a two-model architecture that decouples these concerns via a natural language interface: a multimodal content model generates structured text summaries, and a text-only policy model classifies them according to policy definitions.
      • An iterative co-training loop refines the content model via GRPO to generate policy-relevant summaries, while text-space augmentation generates adversarial summary variants (an augmentation path not achievable on raw multimedia), enabling few-shot policy guidance.
    • EN Key Points:
      • arXiv:2609.22094v1 Announce Type: new
      • Abstract: Content moderation systems traditionally entangle multimodal understanding with policy-specific classification, requiring full pipeline retraining for…
      • We propose Summarize-Judge-Refine (SJR), a two-model architecture that decouples these concerns via a natural language interface: a multimodal Content Model pro…
      • An iterative co-training loop refines the Content Model via GRPO to produce policy-relevant summaries, while text-space augmentation generates adversarial summa…

  • AI-inferred expressed well-being and collective-action discourse in climate-change campaigns on X

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22096v1 Announce Type: new.
      • Abstract: Climate campaigns are often evaluated through attention and mobilization, but less is known about the well-being language that accompanies them.
      • Whether campaign periods alter positive affect and hope, and whether happiness aligns with action language, remains unresolved.
      • We analysed 364,118 public Twitter/X posts from Earth Day, Earth Hour, Global Climate Action Day and World Environment Day in 19 occurrence-years, using 30-day pre-event, event, and post-event windows.
    • EN Highlights:
      • arXiv:2609.22096v1 Announce Type: new
      • Abstract: Climate campaigns are often evaluated through attention and mobilization, but less is known about the well-being language that accompanies them
      • Whether campaign periods alter positive affect and hope, and whether happiness aligns with action language, remains unresolved
      • We analysed 364,118 public Twitter/X posts from Earth Day, Earth Hour, Global Climate Action Day and World Environment Day in 19 occurrence-years, using 30-day…

  • Token Signatures of Code: Comparing Coding Behaviors Across Large Language Models

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22097v1 Announce Type: new.
      • Abstract: The evaluation of large language models (LLMs) on coding tasks has primarily focused on performance metrics such as pass@k.
      • As LLMs continue to advance, many models now meet baseline performance requirements, reducing the discriminative power of performance-based evaluation alone.
      • Yet a key question remains largely unexplored: how do LLMs differ in their coding behavior?
    • EN Highlights:
      • arXiv:2609.22097v1 Announce Type: new
      • Abstract: The evaluation of large language models (LLMs) on coding tasks has primarily focused on performance metrics such as pass @projects/huanxi-teams-ultra/requirements/BACKLOG.json
      • As LLMs continue to advance, many models now meet baseline performance requirements, reducing the discriminative power of performance-based evaluation alone
      • Yet a key question remains largely unexplored: how do LLMs differ in their coding behavior

  • TreeSpark: Calibrated, Load-Adaptive Draft Trees for Semi-Autoregressive Speculative Decoding

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22098v1 Announce Type: new.
      • Abstract: Speculative decoding accelerates language model inference by having an inexpensive drafter propose tokens that the target model validates in parallel.
      • Recent block drafters make drafting nearly free: a single backbone pass emits an entire block of draft tokens.

  • Draft trees promise further gains—verifying several alternative continuations in one target—but existing structures rank candidates by per-position edges, ignoring the parents from which candidates extend. Consequently, in semi-autoregressive drafters, wider trees mostly add misranked nodes, and fixed-size trees ignore how much speculation can be supported per decoding round and per serving load.

    • EN Highlights:
      • arXiv:2609.22098v1 Announce Type: new
      • Abstract: Speculative decoding accelerates language-model inference by letting a cheap drafter propose tokens that the target model verifies in parallel
      • Recent block drafters make drafting nearly free: a single backbone pass emits an entire block of draft tokens
      • Draft trees promise a further gain – several alternative continuations verified in one target forward – but existing constructions rank candidates by per-posi…

  • A framework for recipe data structure with applications for culinary and nutritional insights

    • Release Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22099v1 Announce Type: new.
      • Abstract: Cooking is a complex process that transforms raw ingredients into delicious and nutritious dishes, yet the recipes that encode this process remain largely free-text; readable by people but not directly computable.
      • Existing recipe collections capture fragments of this information, but no shared representation links a recipe’s structured ingredient composition, its geo-cultural origin, and its nutritional profile in a single queryable schema.
      • We address this representation gap by formalizing a framework for recipe data structure that decomposes each recipe into typed ingredient entities, grounds those entities in a reference nutritional database, and annotates them with geo-cultural and dietary context.
    • EN Highlights:
      • arXiv:2609.22099v1 Announce Type: new
      • Abstract: Cooking is a complex process that transforms raw ingredients into delicious and nutritious dishes, yet the recipes that encode this process remain lar…
      • Existing recipe collections capture fragments of this information, but no shared representation links a recipe’s structured ingredient composition, its geo-cult…
      • We address this representation gap by formalizing a framework for recipe data structure that decomposes each recipe into typed ingredient entities, grounds thos…

  • AdaMem: Adaptive Memory Token Allocation for Soft Compression in Retrieval-Augmented Generation

    • Release Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22100v1 Announce Type: new.
      • Abstract: Retrieval-Augmented Generation (RAG) improves language models with retrieved evidence, but processing many long passages is costly and can introduce distracting information.
      • Soft compression addresses this challenge by encoding passages into a compact sequence of continuous memory embeddings prior to generation.
      • However, existing methods typically allocate the same number of memory embeddings for each retained passage, without considering its query-specific relevance.
    • EN Highlights:

  • arXiv:2609.22100v1 Announce Type: new

  • Abstract: Retrieval-augmented generation (RAG) improves language models with retrieved evidence, but processing many long passages is costly and can introduce d…

  • Soft compression addresses this challenge by encoding passages as compact sequences of continuous memory embeddings before generation

  • However, existing methods typically assign each retained passage an identical number of memory embeddings, irrespective of its query-specific relevance

  • Context Poisoning as Extreme-Value Attention Interference in Long-Context Language Models

    • Release Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22101v1 Announce Type: new.
      • Abstract: Large language models can process increasingly long prompts, but their ability to locate and use decisive evidence may be diminished as irrelevant or confusing context is added.
      • We formulate this phenomenon, which we call context poisoning, as extreme-value interference in attention: the score for decisive evidence is upper-bounded, while the maximum score of effective distractors grows with their number.
      • Under a softmax retrieval abstraction, we derive a finite-sample upper bound showing that maintaining a fixed accuracy target above the base rate requires the evidence margin to scale as $\Omega(\sqrt{\log N})$, where N represents the effective distractor count, not necessarily the original context length.
    • EN Key Points:
      • arXiv:2609.22101v1 Announce Type: new
      • Abstract: Large language models can process increasingly long prompts, yet their ability to locate and use decisive evidence may degrade as irrelevant or confus…
      • We formulate this phenomenon, which we call context poisoning, as extreme-value interference in attention: the decisive-evidence score is upper-bounded, while t…
      • Under a softmax retrieval abstraction, we derive a finite-sample upper bound showing that maintaining a fixed accuracy target above base rate requires the evide…

  • DeepInstructor: An Agentic AI Instructor for Experience-Driven Idea Evaluation

    • Release Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22104v1 Announce Type: new.
      • Abstract: With advancements in automated scientific discovery, Large Language Models (LLMs) can now generate research ideas at an unprecedented scale, shifting the bottleneck from idea generation to idea evaluation.
      • Existing evaluators primarily rely on parameterized LLM knowledge or unstructured retrieval, producing judgments that lack the experience-based reasoning used by human instructors.
      • To address this, we propose DeepInstructor, an agentic framework that formulates idea evaluation as reasoning over structured academic experience.
    • EN Key Points:
      • arXiv:2609.22104v1 Announce Type: new

  • Abstract: As automated scientific discovery advances, Large Language Models (LLMs) can now generate research ideas at an unprecedented scale, shifting the bottl…

  • Existing evaluators mainly rely on parametric LLM knowledge or unstructured retrieval, producing judgments that lack the experience-grounded reasoning used by h…

  • To address this, we propose DeepInstructor, an agentic framework that formulates idea evaluation as reasoning over structured scholarly experience

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

  • PRQuant: Permutation Residual Quantization for Low-Overhead Inference

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22106v1 Announcement Type: New.
      • Abstract: The accuracy of low-bit quantization for linear layers is often determined by a small number of outliers.
      • While existing methods (e.g., smoothing, rotation, or residual-based approaches) can mitigate this issue, they often introduce new accuracy bottlenecks for the weights.
      • Furthermore, most of these techniques are implemented as online methods, which can lead to significant execution overhead.
    • EN Key Points:
      • arXiv:2609.22106v1 Announce Type: new
      • Abstract: Accuracy of Low-bit quantization of linear layers is often dominated by a small number of outliers
      • Although existing methods, such as smoothing, rotation, or residual-based approaches, may mitigate this problem, they often introduce new accuracy bottlenecks t…
      • Besides, most of these techniques are implemented as online approaches, which can result in heavy execution overheads

  • Generalized Multimodal Foundation Model

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22107v1 Announcement Type: New.
      • Abstract: Utilizing multimodal data for prediction is widely applied in various scenarios.
      • Existing multimodal fusion models, once deployed, can only handle predefined modalities (e.g., vision, text, and audio) and single tasks, making it difficult to quickly adapt to new downstream applications.
      • This raises a natural yet rather radical question: does a universal multimodal fusion model exist that can be applied to any combination of modalities and any prediction task.
    • EN Key Points:
      • arXiv:2609.22107v1 Announce Type: new
      • Abstract: Making prediction with multimodal data is widely used in diverse scenarios
      • Existing multimodal fusion models, once deployed, can only handle predefined modalities (e.g., vision, text and audio) and single tasks, making it difficult to…

  • Therefore, a natural yet rather aggressive question arises, whether there exists a general multimodal fusion model that can be applied to arbitrary modality com…

  • Correcting Learning-based Perception for Safety

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22108v1 Announcement Type: New.
      • Abstract: Learning-enabled perception is important in many autonomous systems.
      • Unlike traditional sensors, the boundary where machine learning perception works or does not work is difficult to characterize.
      • Incorrect perception can lead to unsafe or overly conservative downstream control actions.
    • EN Key Points:
      • arXiv:2609.22108v1 Announce Type: new
      • Abstract: Learning-enabled perception is important in many autonomous systems
      • Unlike traditional sensors, the boundary where ML perception does or does not work is poorly characterized
      • Incorrect perception can lead to unsafe or overtly conservative downstream control actions

  • A Shared Learning Rate Is Not a Neutral Control in Selective On-Policy Distillation

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22109v1 Announcement Type: New.
      • Abstract: Selective on-policy distillation trains a student only at the token positions where the selector scores highest, and the literature compares selectors under a single shared learning rate (chosen as a neutral control).
      • Under LoRA on GSM8K (Qwen2.5-1.5B student, 7B teacher), across an 8x learning rate grid, dense supervision is statistically flat (swing 1.8 pp, p=0.26), whereas each selective arm’s performance varies: a random 5% subset by 5.4 pp, the total variation selector by 6.7 pp, and the teachability selector by up to 17.7 pp.
      • Therefore, the verdict between dense and selective supervision is 10.1 pp at lr=1e-4 but 5.1 pp at 5e-5, a 2.0x difference determined by a parameter the protocol treats as incidental, and two of the six pairwise significance calls between selectors flip between adjacent rates without any rank reversal.
    • EN Key Points:
      • arXiv:2609.22109v1 Announce Type: new
      • Abstract: Selective on-policy distillation trains a student only at the token positions a selector scores highest, and the literature compares selectors under a…
      • We show it is not
      • Under LoRA on GSM8K (Qwen2.5-1.5B student, 7B teacher), across an 8x learning-rate grid, dense supervision is statistically flat (swing 1.8 pp, p=0.26) while ev…

  • Toward Fairness in Machine Learning Models for Predicting Treatment Retention and Premature Discontinuation in Medication for Opioid Use Disorder

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22113v1 Announcement Type: New.

  • Abstract: Persistent low retention and completion rates in medications for opioid use disorder (MOUD) have driven the use of machine learning (ML) models to predict retention and identify patients at risk for premature discontinuation.

  • However, the fairness of these models across patient populations remains largely unexplored, raising concerns about their application in treatment decision support.

  • This study systematically assesses algorithmic fairness in ML models for predicting MOUD retention and premature discontinuation and investigates the effectiveness of bias mitigation techniques.

  • EN Highlights:

    • arXiv:2609.22113v1 Announce Type: new
    • Abstract: Persistent low retention and completion rates in medications for opioid use disorder (MOUD) have driven the use of machine learning (ML) models to pre…
    • However, the fairness of these models across patient populations remains largely unexplored, raising concerns about their application in treatment decision supp…
    • This study systematically assesses algorithmic fairness in ML models for predicting MOUD retention and premature discontinuation and investigates the effectiven…

  • ZoAQ: Adaptive Zeroth-Order Querying via Query-Reuse Coupling

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22115v1 Announce Type: new.
      • Abstract: Zeroth-order optimization (ZOO) estimates updates from function evaluations, making perturbation queries a primary cost.
      • Fixed budgets spend the same number of queries at every step, while adaptive controllers may offset their savings by using additional oracle calls to test the reliability of the estimator.
      • We introduce ZoAQ, an adaptive ZOO method built around query reuse.
    • EN Highlights:
      • arXiv:2609.22115v1 Announce Type: new
      • Abstract: Zeroth-order optimization (ZOO) estimates updates from function evaluations, making perturbation queries a primary cost
      • Fixed budgets spend the same number of queries at every step, while adaptive controllers may offset their savings by using additional oracle calls to test estim…
      • We introduce ZoAQ, an adaptive ZOO method built around query reuse

  • LE4Mob: Towards Inductive, Distance-Aware and General-Purpose Location Embedding for Human Mobility Modelling

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22117v1 Announce Type: new.
      • Abstract: Location representations provide mobility models with essential information about spatial locations, functional characteristics, and the relationships between places.
      • However, existing embeddings often rely on mobility observations, are unable to represent unseen locations, and have weak constraints for preserving geographical distance.
      • This limits their reuse across datasets and mobility tasks.
    • EN Highlights:
      • arXiv:2609.22117v1 Announce Type: new

  • Abstract: Location representations provide mobility models with fundamental information about the spatial position, functional characteristics, and relationship…

  • However, existing embeddings are often dependent on mobility observations, unable to represent unseen locations, and weakly constrained to retain geographic dis…

  • This limits their reuse across datasets and mobility tasks

  • Success Leaves Detours: Learning Executable Walkthroughs for Long-Horizon Agents

    • Publish Date: 2026-09-22 12:00 Beijing Time
    • Abstract:
      • arXiv:2609.22120v1 Announce Type: new.
      • Abstract: Test-time self-evolving agents improve by reusing past experience, yet sparse-reward trajectories contain failures, loops, and detours, while summaries often ignore the state conditions and operational dependencies required for execution.
      • We study executable walkthrough induction from sparse-reward trajectories: extracting compact, state-conditioned, and verifiable procedures.
      • Our main observation is that delayed credit identifies actions associated with progress but cannot determine whether they produce facts required by subsequent actions.
    • EN Key Points:
      • arXiv:2609.22120v1 Announce Type: new
      • Abstract: Test-time self-evolving agents improve by reusing past experience, yet sparse-reward trajectories contain failures, loops, and detours, while summarie…
      • We study executable Walkthrough induction from sparse-reward trajectories: extracting compact, state-conditioned, and verifiable procedures
      • Our key observation is that delayed credit identifies actions associated with progress but cannot determine whether they produce facts required by later actions

  • Modelling daily activity patterns from mobile phone location data via deep representation learning

    • Publish Date: 2026-09-22 12:00 Beijing Time
    • Abstract:
      • arXiv:2609.22121v1 Announce Type: new.
      • Abstract: Passively collected mobile phone location data provide large-scale, longitudinal observations of human mobility but do not directly reveal activity purpose.
      • Functional characteristics of visited locations provide useful contextual information, but their relationship to activity purpose remains uncertain, especially in mixed-use urban environments.
      • We conceptualize activity pattern mining as an integrated process of representation, clustering, and explanation, and propose the Activity Chain Encoder (ACE) for the representation phase.
    • EN Key Points:
      • arXiv:2609.22121v1 Announce Type: new
      • Abstract: Passively collected mobile phone location data provide large-scale, longitudinal observations of human mobility but do not directly reveal activity pu…

  • Rank Portability Does Not Imply Feasibility Portability: Target-Specific Evaluation of Joint Hardware Constraints

    • Publication Time: 2026-09-22 12:00 Beijing Time
    • Abstract: - arXiv:2609.22122v1 Announcement Type: New.
      • Abstract: Cross-device hardware evaluation often assumes that if architecture rankings transfer across devices, a proxy device can support target-side model selection.
      • We stress-test this assumption to validate the joint latency-energy feasibility for two public architecture families.
      • On NAS-Bench-201, cross-device rank correlations are moderate, while target-comparable feasible-set overlap remains incomplete.
    • EN Key Points:
      • arXiv:2609.22122v1 Announce Type: new
      • Abstract: Cross-device hardware evaluation often assumes that if architecture rankings transfer across devices, a proxy device can support target-side model sel…
      • We stress-test this assumption for joint latency-energy feasibility across two public architecture families
      • On NAS-Bench-201, cross-device rank correlations are moderate, while target-comparable feasible-set overlap remains incomplete