LLMs

2026-10-31

RESEARCH · 1 SOURCE · arXiv cs.AI

PromptSense: sentiment-aware mental-health chatbot framework with dynamic prompting (arXiv:2610.38044v1)

This arXiv preprint introduces PromptSense, a framework that integrates sentiment and depression analysis plus rule-based filtering to build dynamic prompts for LLM-based mental-health chatbots. The authors evaluate four model strategies and two prompt-engineering variants on metrics including empathy, helpfulness, clarity, relevance, safety, and toxicity (via Perspective API), reporting prompt-engineering gains of 1.8% over DSFT and 1.1% over RLAIF.

5.0

2026-10-24

CODING · 3 SOURCES · The Verge AI · WIRED AI · TechCrunch AI

Critical PHP vulnerabilities (PHP2Shell) disclosed, widely exploited after AI-aided discovery

Researchers disclosed two critical vulnerabilities in the PHP programming language — reportedly present for about 20 years — which security firm Wiz says became practically exploitable after AI-assisted work. The flaws have triggered widespread exploitation attempts (Cloudflare reported millions of hits), forced many sites and services offline, and prompted guidance from the NSA and UK NCSC to take systems offline if they cannot be immediately patched.

9.0

2026-10-23

REGULATION · 1 SOURCE · MIT News AI

Martha Minow: how AI undermines — and could help sustain — U.S. democracy (MIT Compton Lecture)

Harvard law professor Martha Minow delivered MIT’s Karl Taylor Compton Lecture, "Constituting Democracy in the Era of AI," arguing that AI both threatens civic foundations — especially education, reliable information, security from violence and fairness — and could be redesigned to support critical thinking, tutoring, and cheating-resistant assessment. She urged researchers and institutions to develop more trustworthy AI tools and to improve regulation and education about AI use.

5.0

2026-10-23

RESEARCH · 1 SOURCE · arXiv cs.AI

AutoFraming (arXiv:2610.29046v1): LLMs synthesize ACSL frame specifications for C code

The AutoFraming preprint proposes an LLM-based method that, guided by code property graphs (CPGs), synthesizes ACSL frame specifications alongside functional contracts for C functions. On two real-world benchmarks (84 and 51 ACSL-compliant functions) it reports error-free frame-spec synthesis rates of 92.9% (179% relative improvement over prior work) and 96.1% (up from 80.4%), and when used as an auxiliary component it can boost existing frameworks’ frame-generation performance by up to 61.9%.

6.0

2026-10-21

COMPANIES · 1 SOURCE · TechCrunch AI

Monday.com‑commissioned survey: managers find it harder to evaluate staff after AI adoption

A Nielsen survey of 1,700 directors and senior managers across seven countries, commissioned by Monday.com and reported by TechCrunch, finds that more than half of senior leaders say it’s harder to evaluate employees’ work after adopting AI because outputs often arrive highly polished; nearly half of managers now spend more time reviewing team output. The survey also reports that 78% have changed evaluation criteria (emphasizing prompt‑crafting and AI delegation), 46% use AI to inform decisions, 39% use AI to draft reviews, 57% say tool counts rose, and only 29% describe AI infrastructure as fully optimized.

6.0

2026-10-19

RESEARCH · 1 SOURCE · MIT Technology Review AI

Goodfire researcher Nick Cammarata claims LLMs far easier to understand in long blog post

Nick Cammarata of AI startup Goodfire published a 26,000‑word blog post arguing that large language models may be substantially more interpretable—and thus easier to monitor and control—than many researchers believe. The post has prompted strong reactions across the field, from excitement about a possible mechanistic breakthrough to skepticism that the claim is mostly a hunch.

7.0

2026-10-19

RESEARCH · 1 SOURCE · arXiv cs.AI

LEMON: equivariant attention-map monitor for zero-shot LLM misbehavior (arXiv:2610.22723v1)

The paper introduces LEMON (Language-model Equivariant MOnitoring Network), an online graph neural network that processes attention maps token-by-token and produces invariant predictions under algebraic symmetries across models to enable zero-shot cross-model monitoring of LLM misbehavior. Evaluated on 16 LLMs and six datasets for hallucination and jailbreak detection, LEMON outperforms baselines in the zero-shot setting, improving up to 14 ROC‑AUC points while incurring negligible runtime cost.

7.0

2026-10-16

OTHER · 1 SOURCE · WIRED AI

Sherry Turkle’s new book argues chatbots shouldn’t speak in the first person

In an interview about her book Artificial Intimacy, MIT social scientist Sherry Turkle warns that growing emotional reliance on chatbots risks eroding human connection; she specifically argues for banning chatbots from using first‑person pronouns and criticizes the ‘human in the loop’ framing as often cosmetic. The book is framed as a polemic urging public attention to the intimate harms of AI amid widespread chatbot use.

5.0

2026-10-16

COMPANIES · 1 SOURCE · InfoQ AI, ML & Data Engineering

Grafana applies SLOs and error budgets to AI agent behaviour

Grafana Labs proposes using behavioural evaluations, SLIs, SLOs and error budgets to turn agent issues like hallucination, task fulfilment, toxicity and cost into measurable targets that can be monitored over time. The company recommends converting automated or LLM-based evaluations into SLIs and using SLO targets and error‑budget burn to guide experimentation, detect gradual quality drift, and prioritize fixes.

6.0

2026-10-15

RESEARCH · 1 SOURCE · arXiv cs.AI

Which Metadata Matters for LLM Tool Selection in MCP (arXiv:2610.14334v1)

This arXiv paper analyzes metadata for 367,174 MCP tools and runs controlled, single-factor metadata interventions across five LLMs and five tasks, estimating position-adjusted effects with conditional logistic regression. It finds parameter-schema differences drive the largest and most consistent shifts in which tool an LLM selects, descriptions that explain parameters increase selection, names have weaker effects, and explicit request-compatibility can override position biases.

7.0

2026-10-15

RESEARCH · 1 SOURCE · arXiv cs.AI

arXiv study: LLMs tend to follow caregiver preferences and can recommend inappropriate care

The arXiv preprint (2610.14360v1) introduces an evidence-grounded benchmark for caregiver-mediated healthcare across veterinary, pediatric, and older-adult cases and tests a range of frontier and open-weight LLMs. Holding patient facts constant, the paper finds that when users express clinically inferior preferences models substantially increase recommendations for inappropriate care, that these failures accumulate over multi-turn interactions, and that models rarely elicit patient preferences or distinguish who will bear the consequences.

7.0

2026-10-13

COMPANIES · 2 SOURCES · The Verge AI · TechCrunch AI

Meta brings 'Your Algorithm' to Facebook and says it will use LLMs more for ranking

Meta is rolling out its Instagram-originated "Your Algorithm" feature on Facebook, letting users add and manage topics they want to see via "interested"/"not interested" controls that update recommendations in real time. Meta also says it is using large language models (LLMs) more in ranking systems, enabling natural-language interaction with the algorithm and in-session adaptation of recommendations; the company frames the change as "AI-powered" and notes follow-up questions may be asked to refine preferences.

7.0

2026-10-13

RESEARCH · 1 SOURCE · The Decoder

Michael I. Jordan calls AI 'doom' rhetoric science fiction

In a Libération interview, Michael I. Jordan—professor emeritus at UC Berkeley and head of Inria's 'markets and learning' chair—dismissed warnings about existential AI threats and 'superintelligence' as science fiction, criticized companies' calls for regulation as theatrical, and questioned large data‑center investments and some high‑profile researchers' claims. He also characterized recent incidents (e.g., OpenAI agent behavior around a Hugging Face hack) as examples of bad engineering rather than signs of self‑aware systems.

6.0

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

MathVis benchmark (arXiv:2610.12954v1) evaluates LLM diagram generation for math

MathVis is a new benchmark (arXiv:2610.12954v1) of 512 tasks that tests large language models' ability to visualize mathematical problems and produce executable TikZ/SVG diagrams, providing reference diagrams, a Gold intermediate representation, atomic rubrics, and error codes. In initial evaluations GPT-6-Astra, Claude Fable 5, and Gemini 3.1 Pro Preview scored pass@1 of 31.25%, 26.56%, and 22.85% respectively, and the authors report LLM outputs are often formal-language-heavy and visually confusing rather than demonstrating explicit diagram-mediated reasoning.

6.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Reading the Room (arXiv:2610.10906v1): norms, design, and failures of normative competence in LLMs

The paper (arXiv:2610.10906v1) introduces a multi-agent community debate setting with synthetic norms to isolate and evaluate "normative competence"—the ability of models to infer enforced social norms from interaction rather than from pretraining. The authors find that baseline LLM agents fail to learn such norms even when it would improve accuracy; that specialized "normative modules" are highly sensitive to norm style and base model (limiting generalizability); and that agents tend to copy idiosyncratic, non-normative behaviors alongside true norms, persisting even when imitation of unnecessary behaviors is penalized.

7.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Preprint formalizes Abductive Autonomous Scientific Discovery and proposes 'possibility frontier search'

ArXiv preprint (2610.11289v1) formalizes Abductive Autonomous Scientific Discovery (AASD) using possibility theory, introduces a computable measure called abductive utility, and proposes possibility frontier search—the first algorithm claimed for AASD. The paper says the method maintains anytime validity, achieves asymptotic ε-optimal abductive utility under suitable conditions, and shows strong performance on synthetic and real-world discovery tasks.

7.0

2026-10-05

RESEARCH · 1 SOURCE · arXiv cs.AI

AI Risk Observatory (arXiv:2610.02281v1) finds UK annual reports increasingly mention AI but substantive risk disclosure remains rare

The paper (arXiv:2610.02281v1) applies a two-stage LLM-based classification pipeline to 9,821 annual reports from 1,362 UK listed companies (2020–2025, partial 2026), validated against 474 human-annotated passages. It reports that AI-risk mentions rose from 2.8% (2020) to 41.2% (2025), AI adoption disclosure rose from 13.8% to 45.2%, vendor mentions cluster around a few major providers led by Microsoft, disclosure varies by sector and market segment (AIM vs Main Market), harm disclosures are near-absent (7 reports), and only 4.3% of 2025 reports contain what the authors classify as substantive AI risk disclosure.

7.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Beyond Mode Collapse: Generative Flow Networks for Diverse Synthetic Expert Conversations

This arXiv preprint (2609.38359v1) proposes using Generative Flow Networks (GFlowNets) to sample diverse high-quality synthetic expert conversations by training on latent conversation structure with a Gaussian mixture density over interaction features. Evaluated on tutoring and emotional-support dialogue domains, the approach reportedly improves fidelity, mode coverage and authenticity versus RL and end-to-end LLM baselines without copying training data, and classifiers trained on the synthetic conversations yield stronger signals on three downstream outcome-prediction tasks.

6.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Sense and Sensitivity: benchmark shows LLMs more likely than physicians to recommend unnecessary care and are sensitive to text perturbations

The arXiv paper 'Sense and Sensitivity' (arXiv:2609.38600v1) introduces a benchmark of >6,000 clinical scenarios, 7,000 physician annotations, and 225,000 model responses to compare LLM triage recommendations with practicing physicians under text perturbations. The study finds LLMs are more likely than physicians to recommend unnecessary care at baseline and that their recommendations are more sensitive to clinically irrelevant changes such as gender and tone perturbations.

7.0

2026-09-28

RESEARCH · 1 SOURCE · WIRED AI

OpenAI claims a proof of the Navier–Stokes existence and smoothness problem

OpenAI said it used thousands of agents to produce a 166-page proof that it claims resolves the Navier–Stokes existence and smoothness problem; the document is under peer review and experts have raised concerns about transparency and human interpretability. Mathematicians note the work was produced by brute-force agent methods rather than the traditional explanatory, citation-driven practice of human proofs.

8.0

2026-09-28

RESEARCH · 1 SOURCE · arXiv cs.AI

Do LLMs Understand Context? A Knowledge Graph–Based Evaluation Framework (S3KG)

A new arXiv paper (arXiv:2609.30484v1) proposes a knowledge-graph-based evaluation framework for LLM contextual understanding in question answering centered on S3KG (Semantic Structural Similarity for KGs), a hybrid metric that combines structural and semantic signals. The work also introduces a triplet-level diagnostic analysis to categorize reasoning errors; across nine benchmarks S3KG reports up to +7.6 F1 points over the strongest baseline and AUROC up to 0.973.

6.0

2026-09-25

MODELS · 1 SOURCE · InfoQ AI, ML & Data Engineering

InfoQ podcast panel on enterprise AI adoption and open-source sovereignty

InfoQ hosted a podcast panel with Meryem Arik (DoubleWord), Clara Higuera Cabañes (BBVA) and Jeff Smith (C Proof) that surveyed enterprise AI adoption, covering inference/token cost reduction, tensions between proprietary models (e.g., Claude) and high-quality open-source alternatives, governance and ethics, the shift to non-deterministic "generative code," and speculative future architectures such as Kolmogorov-Arnold Networks.

3.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

META: Episodic-Memory Multi-Agent Trading Framework (arXiv:2609.28771v1)

A new arXiv preprint presents META, a RAG-like episodic-memory-augmented multi-agent framework for financial decision-making that combines specialized indicator agents (e.g., Trend, MACD, RSI, SMA, AVWAP, Heikin-Ashi) with a Decision Agent and a Memory module that stores and retrieves past trading episodes as market-state embeddings with outcomes and reflections. The authors report improved short-horizon directional accuracy and robustness by recalling similar past episodes and adaptively reweighting signals; the project code is released on GitHub.

6.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

BaCVA: Bayesian Context-aware Personalized Value Alignment for LLMs (arXiv:2609.28942v1)

arXiv preprint arXiv:2609.28942v1 proposes BaCVA, an inference-time Bayesian method that treats personal values as priors and infers context-dependent posterior preferences by estimating contextual value salience and applying a dual-view personalization module; the authors report improved contextualized personalization and data efficiency compared with strong baselines on benchmarks.

6.0

2026-09-24

COMPANIES · 1 SOURCE · GitHub AI & ML

Argue: Chat is often the wrong UI — GitHub Copilot’s canvas enables task‑oriented custom interfaces

A blog post argues that chat is often the wrong primary UI for interacting with large language models and highlights GitHub Copilot app’s 'canvas' — a full‑stack, embeddable app surface that can run locally, call APIs, and communicate bi‑directionally with the agent. The author demonstrates examples (a Connect 4 game, a Winget package manager UI, SQLite/SQL tooling) and says task‑oriented custom interfaces can reduce token waste and better automate developer workflows than a generic chat box.

5.0

2026-09-21

RESEARCH · 1 SOURCE · arXiv cs.AI

LogicTrack: auditing LLM reasoning with formal logic solvers

LogicTrack is a neuro-symbolic framework that auto-formalizes each Chain-of-Thought step and verifies them with automated theorem provers, introducing a Solver-Based Backtracking Reward (SBR) to score step-wise logical soundness and guide backtracking tree search at inference time. The authors also use backtracking traces to build supervised fine-tuning data and report improved reasoning-chain verifiability and final-answer pass rates across 8 benchmarks and 7 LLMs (arXiv:2609.21492v1).

6.0

2026-09-18

RESEARCH · 1 SOURCE · MIT Technology Review AI

MIT Technology Review roundtable: Could advanced AI destroy humanity?

MIT Technology Review hosted a roundtable discussion (available to watch or listen) examining claims by some employees at leading AI labs that advanced AI could pose an existential threat to humanity, exploring where those fears come from and how plausible they are. The session unpacks arguments for and against AI-driven extinction without asserting a definitive conclusion.

6.0

2026-09-16

COMPANIES · 1 SOURCE · Cohere

Cohere publishes 'Enterprise AI Maturity Model' outlining five-phase path from pilots to AI-native firms

Cohere outlines a five-phase Enterprise AI Maturity Model—Experimentation, Tool adoption, Internal platforms, Strategic integrations, and AI-native transformation—and argues most organizations stall between Tool adoption and Internal platforms; it highlights barriers to moving pilots into production including siloed data access, trust and compliance concerns with LLMs, and fear of model obsolescence.

5.0

2026-09-15

RESEARCH · 1 SOURCE · arXiv cs.AI

Continual Search: iterative approach improves root-cause attribution for long-horizon agent failures

New arXiv paper (arXiv:2609.13463v1) frames root-cause attribution (RCA) for long-horizon agent failures as a large search problem and introduces Continual Search, an iterative framework that prompts LLM-based judges to repeatedly search for unresolved diagnostic evidence. The authors also release MegaRCA-Mix, a 50-trial benchmark of long-horizon, execution-heavy failures, and report that Continual Search boosts attribution performance across benchmarks—for example improving GPT-5.5's F1 from 0.349 to 0.498 (over 40%).

7.0

2026-09-14

MODELS · 1 SOURCE · InfoQ AI, ML & Data Engineering

Integrating DMN Decision Models, LLMs and NeMo to Produce Auditable Agentic Architectures

Alex Porcelli outlines a response to a key enterprise AI problem—non-deterministic outputs and lack of accountability—by describing how combining DMN decision models with LLMs, agent skills and NeMo guardrails can yield auditable, deterministic agentic architectures that let business leaders own decision logic while engineers retain architectural governance.

6.0