Large language models

Discovery timeline

Items with an evidence label, oldest first (labels: what they mean). Unlabelled items (before the labels began) are listed below by date.

1. 21 OctPreprintWhen to Stop Asking: Factor Coverage Bias and LLM Determinability Judgments with Underspecified…
2. 21 OctPreprintReliable Before Profitable: QuantHarness for Governing Financial Strategy Development with LLM…
3. 21 OctPreprintStrucChunk: Structure-Aware Retrieval for Legal Document RAG
4. 21 OctPreprintLLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal…
5. 21 OctPreprintCan It Actually Work? Synthesizing Functional CAD Mechanisms
6. 21 OctPreprintMoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health…
7. 21 OctPreprintSPLASH: Contrastive Audit Pressure for Detecting and Localizing Security Hallucinations in…
8. 21 OctPreprintSARP: An On-SSD Index for KVCache with Fast Prefix Matching and Recovery
9. 21 OctPreprintFake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
10. 21 OctPreprintSpanGuard: Decoupling Instruction Localization from Authorization for Lightweight IPI Defense
11. 21 OctPreprintTowards More Granular Reasoning: Mapping the DSA Transparency Database Taxonomy to Platform Terms…
12. 21 OctPreprintGame of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams

21 October 2026

• When to Stop Asking: Factor Coverage Bias and LLM Determinability Judgments with Underspecified Information in Medicine

  Clinical decision-making often begins with underspecified information and requires not only gathering missing information but also knowing when to stop.

  In the issue · 04 AI Healthcare · arXiv 2610.25819 · Evidence: Preprint

• Reliable Before Profitable: QuantHarness for Governing Financial Strategy Development with LLM Agents

  LLM agents can turn natural-language financial research requests into executable and backtested strategies, but a successful run may still implement the wrong rule, use information unavailable at…

  In the issue · 05 AI Finance · arXiv 2610.26101 · Evidence: Preprint

• StrucChunk: Structure-Aware Retrieval for Legal Document RAG

  Retrieval-Augmented Generation (RAG) can ground LLM outputs in domain-specific knowledge, but legal documents are hard: standard chunking fragments hierarchical argument structure, and generic…

  In the issue · 06 AI Law & Investigation · arXiv 2610.26313 · Evidence: Preprint

• LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context

  Confidentiality requirements increasingly push legal professionals toward small, on-device LLMs that keep case data in-house, used for everyday tasks such as summarization, translation, and reformulation.

  In the issue · 06 AI Law & Investigation · arXiv 2610.26211 · Evidence: Preprint

• Can It Actually Work? Synthesizing Functional CAD Mechanisms

  While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions.

  In the issue · 10 AI Engineering · arXiv 2610.26861 · Evidence: Preprint

• MoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health Coaching

  Despite rapid advances in large language models (LLMs), their deployment in the sports and health domain remains high-risk due to the absence of domain-specific evaluation benchmarks.

  In the issue · 20 AI Sports & Fitness · arXiv 2610.26185 · Evidence: Preprint

• Towards understanding and monitorability of reasoning LLMs

  An editorial on why the chains of reasoning produced by large language models need to be understood and monitored.

  In the issue · 30 AI Research & Innovation · Nature Machine Intelligence

• Astrolab's upcoming rover will take LLMs to the Moon

  The 480-kilogram FLIP (FLEX Lunar Innovation Platform) is the first in a series of vehicles NASA has contracted from Astrolab.

  In the issue · 31 AI Space & Cosmos · MIT Technology Review

• SPLASH: Contrastive Audit Pressure for Detecting and Localizing Security Hallucinations in Generated Code

  Code generated by large language models often passes every functional test while remaining insecure.

  In the issue · 38 AI Coding & Developer Tools · arXiv 2610.26001 · Evidence: Preprint

• SARP: An On-SSD Index for KVCache with Fast Prefix Matching and Recovery

  Distributed LLM serving relies on a global index to locate reusable key-value cache (KVCache) blocks through prefix matching.

  In the issue · 39 AI Cloud & Data Centres · arXiv 2610.26849 · Evidence: Preprint

• On-Device AI Builders Hackathon

  Local-LLM building on Snapdragon AI PCs.

  In the issue · 41 AI Consumer Devices & Wearables · 7–8 Dec 2026

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Large language model (LLM) agents now trade real capital using news as evidence, yet they cannot reliably tell fabricated headlines from real news.

  In the issue · 42 AI Cybersecurity · arXiv 2610.26438 · Evidence: Preprint

• SpanGuard: Decoupling Instruction Localization from Authorization for Lightweight IPI Defense

  LLM-based guardrails are effective against indirect prompt injection (IPI), but their performance depends heavily on model capability: as the guardrail model becomes smaller or weaker, detection degrades substantially.

  In the issue · 42 AI Cybersecurity · arXiv 2610.26013 · Evidence: Preprint

• Towards More Granular Reasoning: Mapping the DSA Transparency Database Taxonomy to Platform Terms and Conditions Using Large Language Models

  Article 24(5) of the Digital Services Act (DSA) requires providers of online platforms to submit every content moderation decision they make, with a corresponding Statement of Reasons (SoR), to the…

  In the issue · 43 AI Governance · arXiv 2610.26226 · Evidence: Preprint

• Game of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams

  Phone scams impose substantial and growing costs on individuals and financial systems worldwide, yet existing countermeasures impose no resource cost on perpetrators.

  In the issue · 44 AI Social Impact · arXiv 2610.26781 · Evidence: Preprint

• The search for consciousness inside LLMs

  The piece describes an experiment in which researchers watched Claude Sonnet 4.5's internal activity while asking it to count to five and 'introspect deeply'.

  In the issue · 49 AI Consciousness · Mint (The Economist)

20 October 2026

• Complementary Vulnerabilities: Auditing and Governing Clinician-LLM Error Overlap

  Health systems are deploying large language models (LLMs) into clinical workflows, often on the assumption that a second automated check can only reduce error.

  In the issue · 04 AI Healthcare · arXiv 2610.25279

• Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework

  Patient-facing medical question answering with large language models (LLMs) is moving from research demonstrations to production deployments, yet evaluation remains anchored on factual correctness.

  In the issue · 04 AI Healthcare · arXiv 2610.24807

• From SEC Filings to Financial Signals: Evaluating LLMs on Stock Market Reactions and Analyst Forecasts

  Form 10-Q MD&A sections combine long-form management narrative with accounting and forward-looking information.

  In the issue · 05 AI Finance · arXiv 2610.24268

• ACE-Trader: Adversarial Co-Evolution for Generator-Critic Trading Systems

  Generator--critic trading systems pair an LLM-based trader that makes trading decisions with a red team agent that evaluates these decisions and provides feedback for revision.

  In the issue · 05 AI Finance · arXiv 2610.25847

• Can AI agents make law? Evaluating AI legal relevance and creativity

  Large language models, the artificial intelligence (AI) systems behind modern chatbots, are trained on vast amounts of legal text, but can they build on it to write law for situations that text never anticipated?

  In the issue · 06 AI Law & Investigation · arXiv 2610.24661

• A Cue's Hidden Price: Auditing Seller Influence on AI Shoppers

  Large language models are increasingly asked to shop on a user's behalf, which lets seller-controlled content such as badges, endorsements, and framing steer what they recommend.

  In the issue · 07 AI Commerce · arXiv 2610.25843

• Benefit-Targeted Routing and Bounded Revision for LLM-Based Evidence Screening in Manufacturing Supplier Discovery

  An evidence-review system must allocate additional model judgments and control their use.

  In the issue · 14 AI Manufacturing · arXiv 2610.25867

• Do LLMs Reproduce the Age Gap in Loneliness? Evaluating Survey Simulation against European Survey Data

  Large language models (LLMs) are often used to simulate survey respondents, but matching survey averages does not establish accurate demographic differences.

  In the issue · 16 AI Geriatric · arXiv 2610.25158

• AI Representations of Mental Health Stigma: Quantifying the Gap Between AI Estimations and Human Perceptions Across Gender Groups Using GPT-Based LLMs

  The increasing use of artificial intelligence (AI) in mental health applications raises concerns about its potential to internalize and perpetuate societal biases.

  In the issue · 17 AI Psychology · arXiv 2610.16042

• Machine--human agreement in marking undergraduate assignments is specific to model, prompt and institution

  Large language models (LLMs) promise to automate one of higher education's most resource-intensive activities: assessing student writing.

  In the issue · 18 AI Education · arXiv 2610.18019

• Language Models are Lazy Poets: Creative Mode Collapse at Scale

  While aligned large language models (LLMs) can successfully complete many creative tasks, their outputs tend to be highly self-similar.

  In the issue · 24 AI Literature · arXiv 2610.24223

• CSWBench: Evaluating Context-Sensitive Code-Switching in Creative Writing

  Large language models (LLMs) have demonstrated multilingual capabilities across diverse language generation and understanding tasks.

  In the issue · 24 AI Literature · arXiv 2610.20272

• Towards Figurative Language Understanding in Moroccan Darija: An Evaluation of Large Language Models

  Figurative language remains a major challenge for Machine Translation (MT) systems, as the intended meaning of such expressions often goes beyond literal translation.

  In the issue · 25 AI Languages & Translation · arXiv 2610.25406

• Maestro: Agentic Formal Proof Development by Coordination

  Lean is increasingly used to verify mathematical proofs and software produced by LLM agents, but writing formal proofs remains hard.

  In the issue · 30 AI Research & Innovation · arXiv 2610.25920

• Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents

  Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions.

  In the issue · 30 AI Research & Innovation · arXiv 2610.25760

• Decomposing Brain-LLM Alignment with Interpretable Semantic Representations

  Large language models (LLMs) predict neural responses during naturalistic language comprehension, but the linguistic information underlying this brain alignment remains difficult to interpret.

  In the issue · 33 AI Neuroscience · arXiv 2610.21508

• Learning Which Device to Call: Online Tool Grounding for LLM Agents in IoT

  The Internet of Things (IoT) gives large language model (LLM) agents access to physical observations, but one requested capability is usually offered by several devices, and the agent's plan does not say which one to…

  In the issue · 41 AI Consumer Devices & Wearables · arXiv 2610.25563

• RAGLeak: Supply-Chain Privacy Backdoors in Retrieval-Augmented Generation

  Retrieval-Augmented Generation (RAG) enables large language models to access external knowledge during inference, but its reliance on external knowledge bases also introduces new attack surfaces.

  In the issue · 42 AI Cybersecurity · arXiv 2610.25937

• SoK: PIAtlas: Decomposing the Unbounded Prompt Injection Attack Space

  Prompt injection is the dominant security risk for large-language-model (LLM) applications and agents, yet the field still lacks a coherent map of how these attacks are constructed.

  In the issue · 42 AI Cybersecurity · arXiv 2610.24999

• CXBench: A Benchmark for Process-Compliant LLM Agents in Regulated Customer-Experience Workflows

  Tool-using language agents increasingly perform regulated customer-service workflows, where a correct final state is insufficient: agents must also provide disclosures, obtain consent, and follow required procedures.

  In the issue · 43 AI Governance · arXiv 2610.25954

• AICTE-ATAL FDP on Responsible AI Governance for LLMs

  Free six-day offline faculty programme, 50 seats.

  In the issue · 43 AI Governance · 16–21 Nov 2026

• Looking Complete Is Not Being Grounded: Evaluating LLM Sitreps Against Human-Authored Situation Reports

  Evaluation for automatic crisis reporting should measure what practitioners need, not what is convenient to compare.

  In the issue · 44 AI Social Impact · arXiv 2610.18762

• Recoverability-Controlled Safety Miscalibration in Large Language Models

  Safety-aligned language models are expected to refuse harmful requests while remaining useful for benign ones.

  In the issue · 45 AI Ethics & Safety · arXiv 2610.22425

• When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight Perturbations

  Harmful fine-tuning can rapidly erode the safety alignment of large language models, creating substantial risks when aligned models are adapted to downstream tasks.

  In the issue · 45 AI Ethics & Safety · arXiv 2610.25683

19 October 2026

• AI Consilium: benchmarking LLM second opinions

  Scores second opinions on 300 NEJM cases for grounding, actionability and safety.

  In the issue · 04 AI Healthcare · arXiv 2610.22298

• Same person, different advice

  Audits LLM financial advisers. In a US panel, 41% of households changed their risk-tolerance answer within two years, which shifts the advised allocation.

  In the issue · 05 AI Finance · arXiv 2610.23060

• LexPilot: legal LLMs with laypeople

  Shows that legal LLMs often answer underspecified lay questions irresponsibly, and benchmarks how proactively they ask clarifying questions in Chinese civil cases.

  In the issue · 06 AI Law & Investigation · arXiv 2610.23141

• How LLMs respond to escalating suicide risk

  Researchers tested ChatGPT, Gemini and Doubao across 12 five-turn scenarios. Timely intervention reached 87.5% when the user's history was gathered into one prompt, against 50% in native multi-turn chat and 41.7% for…

  In the issue · 17 AI Psychology · Frontiers in Psychiatry

• Occupational gender bias across the LLM lifecycle

  Tracks how gender disparities in job associations change from pretraining to post-training.

  In the issue · 19 AI Work & Employment · arXiv 2610.23587

• The cognitive cost of LLM writing

  Compares reading times for human and LLM text to measure reading effort.

  In the issue · 24 AI Literature · arXiv 2610.23715

• Revieve launches its Beauty LLM Suite

  Specialised AI covering skin, ingredients and beauty shopping.

  In the issue · 26 AI Fashion & Beauty · Revieve

• Causal value of marketing labels in LLM ads

  A deconfounder approach that values the expertise behind AI-generated ads.

  In the issue · 28 AI Marketing & Advertising · arXiv 2610.23025

• CriticaLLM: an LLM game master

  An autonomous game master for tabletop role-playing.

  In the issue · 29 AI Gaming & Immersive Worlds · arXiv 2610.21525

• Deep LLM layers expose passage context in EEG

  19-channel EEG recorded during silent reading aligns with deep LLM layers.

  In the issue · 33 AI Neuroscience · arXiv 2610.23638

• VISTA: steering LLMs with latent user values

  Personalises a model from a few survey answers.

  In the issue · 37 AI Foundation Models & Assistants · arXiv 2610.23702

• PocketTSLM: on-device time-series LLM

  Privacy-preserving reasoning over smartphone and wearable sensor data.

  In the issue · 41 AI Consumer Devices & Wearables · arXiv 2610.21308

• Cultural (in)competence of LLMs

  Tests how well LLMs detect culture-specific stereotypes.

  In the issue · 44 AI Social Impact · arXiv 2610.22545

18 October 2026

• Citation-grounded earnings-call benchmark

  A benchmark that tests whether LLM analysis of earnings-call transcripts is trustworthy and properly cited.

  In the issue · 05 AI Finance · arXiv

• No Hindsight for LLM Fact-Checkers

  Measures the channels through which information leaks into tests of misinformation detection, which can make LLM fact-checkers look better than they are.

  In the issue · 08 AI Journalism · arXiv

• How robust is multimodal claim verification to LLM rewriting?

  Tests whether claim-verification systems still work when claims are rewritten by LLMs.

  In the issue · 08 AI Journalism · arXiv

• DistressBench

  A benchmark that evaluates crisis support from large language models across several dimensions.

  In the issue · 17 AI Psychology · arXiv

• Cultural Awareness is Represented but Not Decoded

  Traces mythological knowledge across 18 open-source LLMs. It finds that cultural awareness is represented inside the models but not decoded in their outputs.

  In the issue · 22 AI Mythology · arXiv

• Improving Diversity in LLM Short Story Generation

  Methods to make LLM-written short stories more diverse.

  In the issue · 24 AI Literature · arXiv

• Can LLMs Tell Stories Out of Order?

  A corpus and benchmark for generating nonlinear narratives.

  In the issue · 24 AI Literature · arXiv

• UTP-Bench: uncertainty-aware travel planning

  A large benchmark testing whether LLM itineraries survive transport delays and crowd fluctuations.

  In the issue · 27 AI Travel & Hospitality · arXiv

• LLM-driven search finds quantum circuits

  Anthropic, Harvard and QuEra show an LLM-driven search discovering state-of-the-art quantum circuits (see 35).

  In the issue · 30 AI Research & Innovation · Anthropic

• Quantum Computing Report on the result

  Independent coverage of the LLM-discovered circuits.

  In the issue · 35 AI Quantum Computing · Quantum Computing Report

• Discovering state-of-the-art quantum circuits with LLMs

  The paper behind the Anthropic–Harvard–QuEra result.

  In the issue · 35 AI Quantum Computing · arXiv

• LLM-jp-4 Technical Report

  Japanese sovereign 8B dense and 32B-A3B MoE models, trained from scratch on about 12 trillion tokens.

  In the issue · 37 AI Foundation Models & Assistants · arXiv

• An Optical KV Cache Network for Disaggregated LLM Serving

  Optical circuit switching to move KV caches between GPUs faster and cut time to first token.

  In the issue · 39 AI Cloud & Data Centres · arXiv

• AI evaluations look unbiased without being fair

  LLMs used as evaluators still respond to race and gender cues while appearing neutral.

  In the issue · 45 AI Ethics & Safety · Yale Insights

• Where Is the 'Self' of a Large Language Model?

  Traces an LLM's 'self' from self-index to self-reference.

  In the issue · 49 AI Consciousness · arXiv