large language model (LLM)

2026-10-26

RESEARCH · 1 SOURCE · arXiv cs.AI

arXiv:2610.32045v1 — Study finds AI-for-Science agent harnesses can degrade LLM safety

The paper evaluates five AI-for-Science agent harnesses across four recent GPT variants on 50 harmful chemistry prompts. It reports that introducing an agent harness can increase harmful compliance under direct prompting and enable adaptive attacks that were ineffective against the base LLM, concluding that the base model's safety is not durable once integrated into a harness and recommending domain-specific safety benchmarks for agents.

7.0

2026-10-07

COMPANIES · 1 SOURCE · TechCrunch AI

Meta rolls out LLM-based tools to detect ads that direct users to child sexual abuse material

Meta says it acted on 33.2 million pieces of child sexual exploitation content on Facebook and Instagram in H1 2026, with more than 97% detected by its systems before user reports. The company announced new AI measures — including a large language model to detect “signposting” (ads that appear benign but point to illegal content), additional automated scans, a red‑teaming AI agent to probe weaknesses, and systems to catch users who return with new accounts — and says it will block destinations and take action against accounts that violate its rules.

8.0

2026-10-07

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

Building a session-ordered Kafka pipeline in Go (InfoQ)

An InfoQ article describes a Go-based design for preserving per-session ordering over Kafka using consistent hashing, a two-level worker hierarchy, and one goroutine per active session. The author explains in-place retries, contiguous watermark commits, and operational hardening (rebalance draining, backpressure, idempotency, dead-letter queues) and reports the system has processed >40M production messages with no observed ordering violations and achieved >100k send-side messages/sec in tests.

5.0

2026-10-07

RESEARCH · 1 SOURCE · arXiv cs.AI

RadOnc-Agent: LLM-orchestrated framework for longitudinal radiotherapy workflows (arXiv:2610.06923v1)

RadOnc-Agent is an agentic AI framework that formalizes radiotherapy into four clinical phases and exposes 26 callable functions via a conversational LLM controller that maps clinical intent to schema-constrained service calls while preserving patient and workflow state. Evaluated on 2,600 single-function requests, 200 synthetic cross-stage scenarios, and 120 workflow instances from 60 de-identified patient records, the system selected the intended function in 98.79% of single-function runs and completed ~96.5–96.7% of cross-stage and real-patient workflows; the authors emphasize these results demonstrate technical feasibility, not clinical correctness or benefit.

7.0

2026-09-21

RESEARCH · 1 SOURCE · arXiv cs.AI

SpecOpt: contact-diff agentic method to optimize binding specificity of existing small molecules (arXiv:2609.21165v1)

This arXiv preprint defines 'specificity optimization' (SpecOpt): making constrained structural changes to existing drugs to increase binding preference for an intended target over measured off-targets while keeping drug-like properties. The authors provide a ChEMBL-derived benchmark and an agentic pipeline that docks compounds to targets and off-targets, encodes residue-aware atom–protein contact differences, asks a large language model to propose modifications, and filters candidates by similarity, ADMET, and docking selectivity; on 915 compounds the mean target–off-target binding gap shifted from −0.72 to +0.47 kcal/mol and 84.8% of compounds improved while mean Tanimoto similarity remained 0.72; ablations show residue-specific contact identities are critical.

7.0

2026-09-18

RESEARCH · 1 SOURCE · arXiv cs.AI

LLM-assisted hierarchical LLM–MARL architecture for agentic low-altitude wireless networks

This arXiv preprint (2609.19538v1) proposes a hierarchical hybrid architecture combining large language models (LLMs) and multi-agent reinforcement learning (MARL) in a dual-loop design: an outer LLM-assisted adaptation loop interprets service requirements and operator intent to reconfigure objectives and resource priorities, while an inner decentralized, parameter-conditioned MARL loop executes policies for heterogeneous unmanned aerial systems in low-altitude wireless networks. A logistics-monitoring case study shows the framework can adapt coordination across services to evolving conditions without retraining MARL policies; the paper also outlines challenges for scalable, trustworthy deployment.

5.0