Skip to content

AI Digest for

This page is the archived edition of one specific day. It does not change any more.

Latest digest Subscribe

Weekly recap: News

17 Aug to 23 Aug 2026 50 articles

AI Models and Research

  • OpenAI pauses frontier model scaling to implement stricter security, monitoring, and alignment safeguards amid cyber-⁠critical capability concerns and previews Private Safety Processing for zero-⁠data-⁠retention frontier models [1, 2, 3, 4, 5].
  • Z.ai releases GLM-⁠5.3, achieving frontier coding performance and emergent cyber capabilities through extended post-⁠training; weights are delayed for security review and now available on API at 1.4/4.4 $ per M tokens [6, 7, 8, 9, 10].
  • Qwen 3.8 27B achieves strong benchmarks, running efficiently on local hardware and scoring 52 on the Artificial Analysis Intelligence Index [11, 12].
  • Anthropic reports no plans to release a stronger internal model ("Model 2") due to rising risks and verification challenges [13].
  • OpenAI’s ChatGPT Computer History feature on macOS logs all clicks/keystrokes in plaintext, raising privacy concerns [14].

LLM Tooling, Agents, and RAG

  • Claude Managed Agents enable ABC Legal to scale a governed fleet of specialized, versioned, and observable agents across 1,100 employees [15].
  • Cloudflare adds MCP traffic detection and Zero Trust controls for AI agent traffic, including Gateway selectors and OAuth support for MCP Portals [16, 17].
  • Needle 2 (45M params) is a tiny, efficient open model for tool calling and structured extraction, outperforming larger models on benchmarks [18].
  • Hex’s DataBench evaluates AI agents on realistic analytics tasks, finding strengths in evidence gathering but weaknesses in open-⁠ended judgment [19].
  • Google Antigravity introduces Custom Agents for scoped, file-⁠based agent configurations to reduce token overhead and improve task focus [20].
  • Agent Safehouse provides a macOS sandbox to restrict local LLM access to only necessary files/integrations [21].
  • Stripe acquires OpenRouter for >$7B, gaining a model-⁠routing platform to strengthen its AI footprint [22, 23, 24].
  • Vercel’s AI SDK adds Code Mode so models can write JavaScript/TypeScript that invokes AI SDK tools [25].
  • GitHub Copilot for JetBrains gains enterprise-⁠managed settings for plugins, MCP server allowlists, OpenTelemetry, and permission modes [26].
  • Replit introduces Free Mode powered by GPT-⁠5.6 Luna to let anyone build working software without token-⁠cost concerns [27, 28].
  • GraphRAG is profiled as a technique to answer cross-⁠document questions by structuring retrieval over knowledge graphs rather than flat corpora [29].

MLOps, DevOps, and Engineering Practices

  • dbt Core v1.12 is GA with UDF improvements, Iceberg catalog simplifications, and an opt-⁠in Rust parser for v2.0 prep [30].
  • Apache Hudi 1.0 indexes accelerate lakehouse point lookups to near-⁠database performance [31].
  • DuckDB v2.0 preview introduces a new SQL parser, storage format, and server-⁠focused features, marking a shift toward server use cases [32, 33, 34].
  • OpenSearch 3.8 improves vector ingestion, radial query throughput, and adds gRPC streaming inference for lower latency [35].
  • Docker Desktop 4.86 beta introduces a unified hypervisor for consistent performance across macOS/Windows/Linux [36].
  • Docker expands supply-⁠chain security with new updates to reduce CVEs amid rising AI-⁠authored code and dependency risks [37].
  • AWS IAM Role Manager auto-⁠creates roles for new workloads, simplifying onboarding for users new to AWS [38].
  • Argo Workflows 4.1 adds OpenTelemetry tracing, GPU/device allocation via Kubernetes DRA, and reduced controller memory usage [39].
  • Mojo is now fully open source under Apache 2.0 with LLVM exceptions, targeting GPUs and AI accelerators [40, 41].
  • GitHub adds token-⁠type-⁠specific credential revocation, enabling finer-⁠grained deauthorization during security incidents [42].

Security and Incident Response

  • Critical prototype pollution vulnerability (CVE-⁠2026-⁠33696) in n8n allows RCE via GSuiteAdmin node schema manipulation; immediate upgrades required [43].
  • RingCentral discloses a breach exposing 1.6M accounts’ PII via a social engineering attack [44].
  • Tailscale traces a 16-⁠year-⁠old SQLite WAL race condition causing corruption, leading to upstream fixes and recovery tooling [45].
  • Snowflake QUERY_HISTORY exports can corrupt INT32 values; casting affected columns prevents overflow and bad analysis [46].
  • OWASP releases its Top 10 CI/CD Security Risks, covering dependency abuse, pipeline poisoning, and access control gaps [47].
  • Mozilla revokes a Firefox signing key after an unencrypted copy was exposed in a private GitHub repo [48].
  • MLflow vulnerability (CVE-⁠2026-⁠64849) under active exploitation: SSRF in versions <3.15.0 enables credential theft via DNS rebinding; patch immediately [49].
  • Endor Labs uncovered 14 critical/high-⁠severity flaws in AI agent platforms (NocoBase, Flowise, Langflow, Dify, Activepieces, Kestra, Airflow), including unauthenticated RCE and sandbox escapes [50].
Sources
  1. Pacing model development in an era of cyber-⁠critical capabilities openai.com
  2. Offering Zero Data Retention for frontier models openai.com
  3. Offering Zero Data Retention for frontier models openai.com
  4. OpenAI Slowed Training Over Cyber Risks openai.com
  5. OpenAI Institutes New Safeguards After Hugging Face Breach techcrunch.com
  6. GLM-⁠5.3: Frontier Coding with Emergent Cyber Capabilities z.ai
  7. Z.ai Delays GLM-⁠5.3 Weights Two Weeks After Cyber Score Beats Mythos 5 implicator.ai
  8. GLM-⁠5.3: How Chinese labs keep stride with the frontier interconnects.ai
  9. GLM-⁠5.3 hits the API at 1.4/4.4 per million tokens venturebeat.com
  10. GLM-⁠5.3: Frontier Coding with Emergent Cyber Capabilities z.ai
  11. Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things simonwillison.net
  12. Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index simonwillison.net
  13. Anthropic sees AI risks rising, no plan to release stronger "Model 2" axios.com
  14. ChatGPT Computer History Logs Every Click and Keystroke on Your Mac hwbusters.com
  15. How ABC Legal turned every employee into a builder with Claude Managed Agents claude
  16. Secure all your internal vibe-⁠coded applications, in one click blog.cloudflare.com
  17. How Cloudflare detects MCP traffic and helps secure it blog.cloudflare.com
  18. Needle github.com
  19. Introducing DataBench hex.tech
  20. Introducing Custom Agents antigravity.google
  21. Agent Safehouse github.com
  22. Stripe Clinches Over $7 Billion Deal to Buy AI Firm OpenRouter bloomberg.com
  23. Stripe Will Reportedly Acquire OpenRouter for Over $7B techcrunch.com
  24. [AINews] Stripe buys OpenRouter for $7B latent.space
  25. Code Mode ai-⁠sdk.dev
  26. Enterprise managed settings in GitHub Copilot for JetBrains github.blog
  27. Replit expands access to software creation with GPT-⁠5.6 Luna openai.com
  28. Replit Introduces Free Mode replit.com
  29. GraphRAG: How AI Answers Questions Hidden Across Many Documents blog.bytebytego.com
  30. dbt Core v1.12 is GA dbt.com
  31. Point Lookups on the Lakehouse: How Hudi Indexes Accelerate Read-⁠Heavy Workloads hudi.apache.org
  32. A Preview of DuckDB v2.0 duckdb.org
  33. DuckDB v2.0: Your Database Deserves a Better Parser duckdb.org
  34. A Preview of DuckDB v2.0 duckdb.org
  35. What’s new in OpenSearch 3.8 opensearch.org
  36. Docker Desktop Gets a Hypervisor of its Own cloudnativenow.com
  37. Make zero CVEs your new default docker.com
  38. How AWS IAM Role Manager Rethinks the Starting Point for IAM Roles aws.amazon.com
  39. Argo Workflows 4.1 medium.com
  40. Mojo🔥 is now open source! modular.com
  41. Mojo🔥 is now open source simonwillison.net
  42. Credential revocation and deauthorization by token type github.blog
  43. From a Schema Name to RCE in n8n simonkoeck.com
  44. RingCentral Data Breach Exposed Info of 1.6M Accounts bleepingcomputer.com
  45. How we tracked down a 16-⁠year-⁠old SQLite bug tailscale.com
  46. Snowflake says this 149 GB query scanned -1.5 GB espresso.ai
  47. OWASP Top 10 CI/CD Security Risks owasp.org
  48. Mozilla Revokes Firefox Signing Key After Unencrypted Copy Lands in GitHub theregister.com
  49. MLflow Bug Actively Exploited to Steal Credentials decipher.sc
  50. Hacking Your Life With AI Can Get You Hacked endorlabs.com

Weekly recap: Papers

17 Aug to 23 Aug 2026 67 papers

LLM Reasoning, Evaluation & Safety

  • A framework for LLM reasoning separates evidence interpretation from decision aggregation using a four-⁠field evidence tuple and identifies count-⁠scale drift as a failure mode misaligning vote rules with posterior ordering when source reliabilities differ [1].
  • A single-⁠pass method enables frozen LLMs to answer and abstain by reconstructing pre-⁠steering residuals, avoiding steering-⁠sufficiency interference and recovering up to 8 AUROC points lost in prior approaches [2].
  • Information satisfaction as a reader-⁠centered evaluation axis shows summarization metrics (ROUGE, BERTScore) fail to capture user-⁠specific utility, with persona/background signals better aligning summaries to individual needs [3].
  • Principle-⁠Bench evaluates LLM-⁠as-⁠judge trustworthiness across accuracy, paraphrase robustness, adversarial robustness, and calibration, revealing no single method dominates all axes [4].
  • Wrong but useful: >40% of wrong-⁠answer messages that change final correctness in multi-⁠agent deliberation provide trajectory value (e.g., useful decompositions) [5].
  • Compliance detectors exhibit rule blindness: deleting, permuting, or substituting governing rules barely changes detection accuracy across activation probes and guard models, including policy-⁠conditioned guards [6].
  • Model hypnosis shows inconspicuous prompt cues (e.g., paraphrases, typos) can systematically control model behavior across families and scales [7].
  • BabelSteering transfers English-⁠derived safety refusal directions to multilingual settings via activation steering, increasing harmful-⁠request refusal across eight languages with marginal utility loss but some over-⁠refusal [8].
  • PCA-⁠guided Activation Scaling (PAS) enables monotonic, bidirectional control over LLM sycophancy (Spearman ρ = +0.92, 15.4% average shift per direction) [9].
  • GEOFlagBench (3,200 webpages) reveals gaps in Generative Engine Optimization (GEO) detection despite strong aggregate F1 (0.880) [10].
  • ConceptGuard benchmarks context-⁠sensitive unlearning for dual-⁠use concepts, addressing gaps in disjoint forget/retain evaluations [11, 12, 13].
  • AI4AI-⁠Bench tests recursive self-⁠improvement by having LLM agents rewrite training algorithms within a fixed compute budget [14, 15, 16].
  • Phantom Gains audits self-⁠improvement claims, exposing measurement artifacts (e.g., batching-⁠induced capability changes) and proposing exact per-⁠problem tests [17, 18, 19].
  • MemTrapBench shows retrieved memories can distort reasoning (e.g., fixation, belief distortion), with all tested memory strategies underperforming no-⁠memory baselines by >10% [20, 21, 22].
  • InsufficiencyBench evaluates LLMs on underspecified legal queries, revealing no model exceeds F2=0.46 in identifying missing material information [23, 24].
  • ContractScrub introduces the first benchmark for legal contract scrubbing, testing error detection (e.g., defined-⁠term misuse) in handcrafted documents [25, 26].

Agentic Systems, Planning & World Models

  • Marionette models evolving 3D world states (e.g., articulated skeletons) for interactive games, delegating geometry/occlusion to a zero-⁠parameter renderer and focusing on appearance, improving long-⁠horizon consistency [27].
  • Twin constructs executable world models from interaction alone, inferring rules and goals to clear 97.8% of ARC-⁠AGI-⁠3 levels, often more efficiently than humans [28].
  • SheetCompass uses hierarchical relation graphs for agentic spreadsheet reasoning, capturing cross-⁠table associations and spatial layouts [29].
  • Neurosymbolic embodied agents combine task-⁠directed visual exploration and constrained symbolic planning (PDDL), achieving >90% success on VirtualHome and ALFWorld with open 4B–27B models [30].
  • BATON addresses long-⁠horizon robot manipulation via agentic subtask exploration with transition-⁠aware memory, reducing multiplicative exploration cost [31].
  • HAF adapts VLA models to humanoid whole-⁠body loco-⁠manipulation via hierarchical action flow and spectral latent RL, avoiding direct fine-⁠tuning [32].
  • Recirculation introduces inference-⁠time recurrence for foundation models, reducing perplexity and boosting accuracy with minimal latency overhead [33].
  • Learning When to Think trains models to adaptively choose reasoning modes (NoThink/Short/Long) via GRPO, with modes emerging without collapse and brief modes outperforming Long on easier problems [34, 35, 36].
  • Break It Down, Pass It On finds subtask-⁠level, text-⁠formatted skills transfer more reliably than task-⁠level or code skills, with specificity/abstractness as key properties [37, 38, 39].

Vision-⁠Language & Multimodal Systems

  • Vision-⁠language models exhibit color bias: green-⁠colored positive words shift sentiment predictions toward positive, causing misclassification of negative content and affecting latent representations [40].
  • Rollplex optimizes GPU utilization for VLM post-⁠training by overlapping prefix computation with rollout decoding, reducing memory overhead (e.g., 165 GiB for Qwen2.5-⁠VL-⁠32B) without breaking on-⁠policy semantics [41].
  • ReWEIGH calibrates token-⁠level ordinal visual evidence in LVLMs by aggregating rank-⁠based signals, reducing hallucinations without training [42].
  • MedUAG introduces a unified understanding-⁠and-⁠generation medical multimodal model, alongside MedUAGCorpus (6M+ instances, 14 imaging modalities) and MedUAGBench (12 generation tasks) [43].
  • GS-⁠VLA enables plug-⁠and-⁠play viewpoint canonicalization for frozen VLA policies via Gaussian Splatting, recovering robustness to camera shifts without retraining (e.g., 90%→10% success drop on LIBERO mitigated) [44].

Efficiency, Optimization & Architecture

  • Attention-⁠Aware Transform Coding (AATC) compresses KV caches via transform coding to minimize attention-⁠aware distortion, outperforming uniform quantization on long-⁠context benchmarks [45].
  • DeaMoE groups MoE experts into departments with shared parameters to accelerate small-⁠batch decoding, addressing memory bottlenecks [46].
  • LP-⁠NAS formulates architecture search as a linear program using validation-⁠loss gradients and training-⁠loss Hessians, enabling faster differentiable NAS [47].
  • Proteus introduces incremental memory activation for long-⁠context modeling, progressively expanding memory to force early compression and reduce interference [48].
  • Relation replaces attention with token-⁠mixing primitives (Full/Flash/Linear Relation), achieving lower validation NLL than MHA at matched scales and 76.4–84.9% of FlashAttention throughput [49, 50, 51].
  • Daedalus-⁠150M is a convolution-⁠attention hybrid optimized for CPU inference, using short convolutions to limit memory growth and outperforming larger baselines on a 5-⁠task benchmark [52, 53, 51].
  • Which Eviction Policy Should an LLM Cache Use? finds LFU dominates across workloads (no policy beats it by >0.041pp), while FIFO and streaming SISO trail by up to 8.67pp [54].

Domain-⁠Specific Applications

  • RecipeNet captures field-⁠level interactions and sequential dependencies in recipe data, outperforming tabular models on recipe tasks [55].
  • CytoBERT is a foundation model for cytometry data, pretrained on 50M+ cells across heterogeneous marker panels [56].
  • G-⁠CARL introduces Patient-⁠oriented Medical Report Interpretation (PMRI) via a grounded, checklist-⁠aligned RL framework, balancing factuality and accessibility [57, 58, 59].
  • Claude (Mythos Preview and Opus 4.8/5) designed protein binders against 15 targets (14 successful, 22–35% binding rate) and analyzed NMR/LC-⁠MS data from raw lab files in 19–23 minutes, matching human analysis on hydrogen counts and purity [60].

Benchmarks & Evaluation

  • InstaBind-⁠Lite formalizes Dense Same-⁠Class Attribute Misbinding (DSCAM) in VLMs with 524 images and 9,580 questions [61].
  • CaliBench evaluates physical calibration of video world models by scoring outcomes in discrete, interpretable spaces (e.g., dice faces) [62].
  • FormalTCS benchmarks end-⁠to-⁠end frontier formal TCS research (175 instances), revealing autoformalization as the sharpest bottleneck (11.5 vs. 28.6 Pass@8 for proving) [63, 64].
  • OenoBench offers a wine-⁠domain knowledge benchmark (3,266 MCQs), with accuracy spans of 53%–84% across 16 frontier models [65, 66].
  • SOP-⁠Bench evaluates AI agents on real-⁠world standard operating procedures (SOPs) from healthcare, logistics, and finance [67].
Sources
  1. Split the Labor: Separating Evidence Interpretation from Decision Aggregation arxiv cs.CL
  2. You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model arxiv cs.CL
  3. Information Satisfaction: A Reader-⁠Centered Axis for Summarization Evaluation arxiv cs.CL
  4. A Four-⁠Axis Trustworthiness Benchmark for LLM-⁠as-⁠Judge in Principle-⁠Based Regulation arxiv cs.CL
  5. Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-⁠Agent Messages arxiv cs.CL
  6. What Do Compliance Detectors Read? An Audit of Activation Probes and Guard Models arxiv cs.AI
  7. Model Hypnosis: Strong control of AI via additive subliminal effects arxiv cs.CL
  8. BabelSteering: Multilingual Safety Alignment via English Steering Vectors arxiv cs.CL
  9. PCA-⁠guided Activation Scaling for Monotonic Bidirectional Control over LLM Sycophancy arxiv cs.CL
  10. GEO-⁠Flag: Detecting and Measuring GEO-⁠Optimized Web Content arxiv cs.LG
  11. ConceptGuard: Benchmarking Context-⁠Sensitive Unlearning in Large Language Models arxiv cs.CL
  12. ConceptGuard: Benchmarking Context-⁠Sensitive Unlearning in Large Language Models arxiv cs.CL
  13. ConceptGuard: Benchmarking Context-⁠Sensitive Unlearning in Large Language Models arxiv cs.CL
  14. AI4AI-⁠Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-⁠Improvement arxiv cs.CL
  15. AI4AI-⁠Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-⁠Improvement arxiv cs.CL
  16. AI4AI-⁠Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-⁠Improvement arxiv cs.CL
  17. Phantom Gains: Auditing Self-⁠Improvement Against a Measured Null arxiv cs.CL
  18. Phantom Gains: Auditing Self-⁠Improvement Against a Measured Null arxiv cs.CL
  19. Phantom Gains: Auditing Self-⁠Improvement Against a Measured Null arxiv cs.CL
  20. MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use arxiv cs.CL
  21. MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use arxiv cs.CL
  22. MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use arxiv cs.CL
  23. InsufficiencyBench: Evaluating LLM legal advice on underspecified user queries arxiv cs.AI
  24. InsufficiencyBench: Evaluating LLM legal advice on underspecified user queries arxiv cs.AI
  25. ContractScrub: A benchmark for final review of legal contracts arxiv cs.CL
  26. ContractScrub: A benchmark for final review of legal contracts arxiv cs.CL
  27. Marionette: Predicting World States, Rendering Geometry, Painting Appearance arxiv cs.AI
  28. Twin: Playing an Unknown Game with a Test-⁠Time Digital Twin arxiv cs.AI
  29. SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning arxiv cs.AI
  30. Neurosymbolic Embodied Agents arxiv cs.CL
  31. Don’t Drop the BATON: Long-⁠Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-⁠aware Memory arxiv cs.AI
  32. HAF: Adapting Generalist VLAs to Humanoid Whole-⁠Body Loco-⁠manipulation via Hierarchical Action Flow and Spectral Latent RL arxiv cs.AI
  33. Recirculation arxiv cs.LG
  34. Learning When to Think: Adaptive Reasoning for Test-⁠Time Compute Allocation arxiv cs.AI
  35. Learning When to Think: Adaptive Reasoning for Test-⁠Time Compute Allocation arxiv cs.AI
  36. Learning When to Think: Adaptive Reasoning for Test-⁠Time Compute Allocation arxiv cs.AI
  37. Break It Down, Pass It On: Cross-⁠Task Skill Transfer in LLM Agents arxiv cs.CL
  38. Break It Down, Pass It On: Cross-⁠Task Skill Transfer in LLM Agents arxiv cs.CL
  39. Break It Down, Pass It On: Cross-⁠Task Skill Transfer in LLM Agents arxiv cs.CL
  40. Seeing Red, Thinking Bad: Color Bias in Vision Language Models arxiv cs.CL
  41. Rollplex: Cross-⁠Phase GPU Spatial Sharing for Vision Language Model Post-⁠Training arxiv cs.LG
  42. ReWEIGH the Evidence: Calibrating Token-⁠Level Ordinal Visual Evidence to Mitigate Hallucinations in Large Vision-⁠Language Models arxiv cs.CL
  43. MedUAG: Unified Understanding and Generation for Medical Multimodal Models arxiv cs.CL
  44. GS-⁠VLA: Plug-⁠and-⁠Play Viewpoint Canonicalization for Frozen VLA Policies via Gaussian Splatting arxiv cs.AI
  45. KV Cache Compression Through the Lens of Transform Coding arxiv cs.CL
  46. DeaMoE: Efficient MoE Structure for Fast Small-⁠Batch Decoding arxiv cs.LG
  47. LP-⁠NAS: Linear Programming-⁠based Neural Architecture Search arxiv cs.LG
  48. Proteus: Incremental Memory Activation for Long-⁠Context Sequence Modeling arxiv cs.CL
  49. Ask Self, Ask Others: Relation Is All You Need arxiv cs.LG
  50. Ask Self, Ask Others: Relation Is All You Need arxiv cs.LG
  51. Daedalus-⁠150M: A Convolution-⁠Attention Hybrid Designed for CPU Inference arxiv cs.CL
  52. Daedalus-⁠150M: A Convolution-⁠Attention Hybrid Designed for CPU Inference arxiv cs.CL
  53. Daedalus-⁠150M: A Convolution-⁠Attention Hybrid Designed for CPU Inference arxiv cs.CL
  54. Which Eviction Policy Should an LLM Cache Use? A Systematic Study Across Workloads, Capacities, and Encoders arxiv cs.LG
  55. RecipeNet: A Hierarchical Transformer for Recipe Data arxiv cs.LG
  56. CytoBERT: A Foundation Model for Cytometry Data arxiv cs.LG
  57. G-⁠CARL: Grounded Checklist-⁠Aligned Reward Learning for Patient-⁠Oriented Medical Report Interpretation arxiv cs.CL
  58. G-⁠CARL: Grounded Checklist-⁠Aligned Reward Learning for Patient-⁠Oriented Medical Report Interpretation arxiv cs.CL
  59. G-⁠CARL: Grounded Checklist-⁠Aligned Reward Learning for Patient-⁠Oriented Medical Report Interpretation arxiv cs.CL
  60. Aug 18, 2026ScienceHow Claude is accelerating protein design and analytical chemistry anthropic research
  61. Diagnosing Dense Same-⁠Class Attribute Misbinding in Large Vision-⁠Language Models arxiv cs.AI
  62. CaliBench: Are the Stochastic Dynamics of Video World Models Physically Calibrated? arxiv cs.LG
  63. FormalTCS: Benchmarking End-⁠to-⁠End Frontier Formal Theoretical Computer Science Research of Large Language Models arxiv cs.CL
  64. FormalTCS: Benchmarking End-⁠to-⁠End Frontier Formal Theoretical Computer Science Research of Large Language Models arxiv cs.CL
  65. OenoBench: A Wine-⁠Domain Benchmark for Knowledge-⁠Grounded Evaluation of Large Language Models arxiv cs.CL
  66. OenoBench: A Wine-⁠Domain Benchmark for Knowledge-⁠Grounded Evaluation of Large Language Models arxiv cs.CL
  67. SOP-⁠Bench: A new benchmark for evaluating AI agents on real business procedures amazon.science