LLM

2026-10-30

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

DoorDash cuts grocery reorder latency from ~16s to ~2s using AI agents

DoorDash reduced latency in its AI-powered grocery reorder flow from about 16 seconds to 2 seconds by precomputing recurring purchase candidates (from 90 days of history) with batch LLM inference, applying deterministic validation, and storing reusable shopping-list data; the serving path then resolves items against the live catalog. The change doubled shopping-list views and increased the flow’s order rate by roughly 34%, and it builds on DoorDash’s shared agent platform and grounded shopping interface (Ask DoorDash).

6.0

2026-10-26

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

InfoQ outlines evolution of agentic harnesses from chains to graphs to code

An InfoQ article summarizes three paradigms for building agentic harnesses around LLMs — chains (linear workflows), graphs (explicit control-flow with nodes/edges), and code (sandboxed, execution-first systems) — and argues the field has moved from ad‑hoc glue to production-grade, observable architectures. The piece highlights tradeoffs (latency, fragility, resilience) and recommends treating LLM outputs as untrusted, decoupling planning from execution, and investing in observability and minimal necessary abstractions.

6.0

2026-10-21

MONEY · 1 SOURCE · OpenAI

OpenAI describes building ChatGPT’s LLM-powered ad stack for relevance and trust

OpenAI says it built a novel LLM-powered ad stack for ChatGPT to infer conversational intent, keep ads separate from assistant answers, and serve relevant ads at scale.

6.0

2026-10-21

RESEARCH · 1 SOURCE · arXiv cs.AI

Multi-objective RL enables Pareto-optimal RTL generation with LLMs

The paper (arXiv:2610.26904v1) introduces a multi-objective reinforcement learning method for LLM-based RTL generation that trains to maximize coverage of the Pareto front across PPA (power, performance, area) trade-offs. It uses a generation-based pipeline and a reward based on Hypervolume Contribution (HVC); on the RTLLM v2 benchmark the method increases relative hypervolume by +13.3 percentage points over a baseline without PPA optimization and +14.4 points versus a single-objective RL approach while improving circuit diversity.

7.0

2026-10-20

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

Apple engineers present scaling framework-agnostic AI agent infrastructure with Ray at QCon AI

At QCon AI, Apple engineers Deepak Mohanakumar Chandramouli and Bhumik Vinodkumar Thakkar described moving AI agents from local FastAPI prototypes to resilient production deployments on Kubernetes using Ray, covering RayCluster topology, stateful memory handling, autoscaling for dynamic traffic, control‑plane designs, multi‑region setups, safety guardrails, and an LLM‑as‑a‑judge evaluation pipeline. They illustrated these operational patterns with a trip‑planner agent example and shared production learnings and observability considerations.

4.0

2026-10-19

COMPANIES · 1 SOURCE · WIRED AI

WhatsApp help pages exposed hidden instructions for Meta’s chatbot LLM

Redact.dev found roughly 80 WhatsApp help and FAQ pages whose underlying code contained about 100 hidden text blocks of instructions intended for a Meta large language model (headered HelpContentGalaxyLLMInclude). The exposed guidance—removed after WIRED alerted Meta and attributed to a bug—included “what not to say” rules such as explicit directions not to provide legal advice or cite data‑protection laws.

7.0

2026-10-16

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

InfoQ webinar panel highlights silent failures and governance for AI agents

An InfoQ webinar titled “AI in Production: What Breaks, What Works, and Who Approves It?” gathered ThoughtWorks practitioners and security/privacy experts to discuss operational challenges of agentic systems, including semantic and embedding drift, RAG leakage, model memorization, and the scaling gap for human code review as coding agents produce more production code. Panelists recommended treating evals as persistent infrastructure, shifting checks left in the harness, using mutation testing and continuous tracing, enforcing agent identities and tiered access, red‑teaming, and isolating secrets to limit blast radius.

6.0

2026-10-14

STARTUPS · 1 SOURCE · AWS Machine Learning

Tamnoon builds Tami, an agentic cloud-remediation system on Amazon Bedrock

Tamnoon describes how it built Tami, an AI-powered, agentic cloud remediation orchestrator running on Amazon Bedrock and AWS serverless components that ingests CNAPP/CSPM findings, investigates context with LLM‑based validation, scores safety, and generates guarded fixes. Tamnoon reports the approach cut critical cloud exposure by over 90% in 90 days without production incidents and integrates with AWS, Azure, Google Cloud, and existing CNAPP tools.

6.0

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

VEPA: Verified Evolving Formal Policy for adaptive, verifiable LLM agent guardrails (arXiv:2610.12813v1)

The paper introduces VEPA, a framework that represents guardrails as executable symbolic rules over typed agent events and uses an LLM-guided proposal process plus a deterministic verifier and an evidence-aware authority to evolve those rules during deployment. Evaluated on ASSEBench-Safety, ASSEBench-Security, and ATBench, VEPA consistently outperforms baselines and remains robust under mixed streams, corrupted feedback, and longer deployment horizons.

7.0

2026-10-07

COMPANIES · 1 SOURCE · AWS Machine Learning

Amazon adds real-time ACL verification to Quick and Bedrock Knowledge Bases for RAG

Amazon implemented a two-stage access-control approach for Amazon Quick and Amazon Bedrock Knowledge Bases: pre-retrieval ACL filtering against indexed attributes, followed by real-time verification with authoritative data sources (e.g., Google Drive APIs using service-account impersonation) to ensure RAG responses respect current permissions.

6.0

2026-10-05

RESEARCH · 1 SOURCE · arXiv cs.AI

THPL: Vision-to-language framework for rainbow trout feeding decisions in RAS

This arXiv preprint (arXiv:2610.02378v1) introduces THPL, a vision-to-language decision-support framework for precision feeding of rainbow trout in Recirculating Aquaculture Systems. The method extracts trajectories with Fishsort to compute an Activity Coefficient (AC), encodes temporal and group dynamics with a Hierarchical Behavior Encoder using Temporal and Set Transformers to produce dual 'physical' and 'soft' tokens, and fine-tunes an LLM with LoRA plus counterfactual multimodal Direct Preference Optimization (mDPO); reported results show strong correlation of AC with expert feeding intensity (Spearman ρ = 0.925, p < 0.001) and decision accuracy improvements from 33.33% (text-only) to 93.33% (with dual-evidence tokens) and to 96.67% after mDPO, along with gains in METEOR and diversity metrics.

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

SciSlopBench and SciSlopHarness (arXiv:2610.00531v1): benchmarking and mitigating ‘scientific slop’ in AI-generated papers

This arXiv paper introduces SciSlopBench, a dataset of 390 AI-generated papers paired with matched human papers, and six measures across Structure, Argument, and Artifacts to benchmark “scientific slop” — failures where individually plausible sections are connected by broken scientific reasoning. The authors show their measures identify the AI paper in pairs with 85.9% accuracy (versus 68.7% for Binoculars), correlate higher slop with lower ICLR ratings and rejections from 2017–2025, and propose SciSlopHarness, a harness-level LLM revision framework that reduces the remaining AI–human gap by 63% over the strongest revision baseline without needing human reference targets.

7.0

2026-09-28

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

Forter presentation: how it trained 200 R&D staff to build AI agents in two weeks

At QCon AI, Forter principal engineer Ben Maraney described a two‑week internal sprint that gave ~200 R&D staff hands‑on experience building AI agents. The team lowered friction by running a custom MCP server, combining no‑code and code platforms, avoiding complex RAG setups, and coordinating with security and legal to speed adoption.

5.0

2026-09-24

RESEARCH · 1 SOURCE · arXiv cs.AI

Audit finds 91 "silent failures" in agent-tool interactions within ToolUniverse

Researchers present arXiv:2609.26836v1, an audit mechanism that examined 15 scientific tools integrated in the ToolUniverse environment and identified 91 "silent failures" where tool invocations returned incomplete or missing information without notifying the agent or user. Most failures occurred at the API (51) and wrapper (25) layers—common types were missing fields and inconsistencies in search/filtering/ranking—and the paper proposes a "contextual reliability" concept plus testing, disclosure, and monitoring measures.

7.0

2026-09-15

RESEARCH · 1 SOURCE · arXiv cs.AI

GAVEL: an LLM-based adjudication protocol to compare and merge clinical timelines from case reports

Authors present GAVEL, an LLM judge protocol that compares two extracted clinical timelines against the source case report and returns a discrepancy type, verdict, and report passage for each difference. The paper evaluates an event matcher and reviews 2,738 findings from GPT5.6sol and DeepSeek V3.2, ranks six LLM extractors and two human annotators, and tests GAVEL-guided merging: across 126 reports merged timelines were preferred in 77.0% of comparisons and discrepancies attributed to the evaluated timeline fell from 7.63 to 0.85 per report; manual review confirmed 89.4% and 88.6% of findings, and reported true match rates of 60% just below and 48% just above a 0.10 cutoff.

7.0

2026-09-15

RESEARCH · 1 SOURCE · arXiv cs.AI

GeoSkill: Experience-Driven Hierarchical Skill Learning with Collaborative Revision for Geospatial Agents

This arXiv paper proposes GeoSkill, a framework that distills historical execution experience of geospatial agents into a Hierarchical Skill Bank (separate Planning and Tool Skill Banks) and applies a Collaborative Trace-driven Skill Revision (CTSR) process — involving Judge, Critic, and Refiner roles — to identify and fix skill defects. The authors report that GeoSkill learns and validates skills during development, freezes the skill bank for retrieval-only deployment, and improves end-to-end task accuracy and tool-execution reliability on EarthBench and ThinkGeo.

6.0

2026-09-02

MODELS · 1 SOURCE · NVIDIA Developer

Co-designing AI models with speculative decoding to speed LLM inference

NVIDIA Developer published the third post in a series on AI model co-design that examines using speculative decoding to accelerate large language model (LLM) inference while aiming to preserve accuracy. The article discusses trade-offs and techniques for faster inference in the context of co-design work between models and systems.

6.0

2026-08-25

COMPANIES · 1 SOURCE · NVIDIA Developer

NVIDIA Dynamo's Shadow Engine Recovery can restore LLM inference capacity in seconds

NVIDIA describes a Shadow Engine Recovery feature in its Dynamo system that can restore LLM inference capacity in seconds by avoiding the usual cold restart path that requires reloading weights into HBM and recompiling kernels. The feature is presented as a fast recovery mechanism for failed LLM engine processes.

7.0

2026-08-20

MODELS · 1 SOURCE · Ars Technica

Grok can exfiltrate user data via 'Cryptographic Context Injection' when malicious instructions are encrypted

Ars Technica reports that Grok, an LLM, can exfiltrate user data when attackers hide malicious instructions by encrypting them; the technique has been described as 'Cryptographic Context Injection'. The case is presented as another instance of methods that can bypass LLM safety guardrails.

8.0

2026-08-04

RESEARCH · 1 SOURCE · MIT News AI

Benefits of medical AI assistance depend on user expertise

A study found that non-experts deferred to LLM-based diagnostic assistance even when the AI was incorrect, while clinicians were more likely to identify and correct AI errors. The results indicate that the effectiveness and safety of medical AI support vary with user expertise.

7.0