AI News Brief Roundup: 2026-11-11 Morning

AI News Brief: Morning roundup, 2026-11-11

Summary

This window was dominated by research on whether AI agents can be trusted: verifying their outputs, tracing where their errors begin, compressing them for cheaper inference, and even proving limits on checking safe programs. Several threads connect these papers. Many propose ways to make agent behavior checkable, from auto-generated verifiers to closed-loop control of model weights, and a recurring finding is that evidence of failure often exists but goes unused. Industry news was lighter, with Armada opening a European AI data-center manufacturing line and Mercor raising at a $20B valuation, so the practical takeaway is that measuring and governing agent behavior is getting as much attention as building more capable models.

1. SteerDistill compresses LLM agents with low-bit quantization while keeping them accurate. A full-precision teacher steers the quantized student to fix failed responses, and the corrections are distilled back into it. It beats the strongest quantized baseline by 1.67 to 2.18 points at W4A16.
2. This paper proves that no finite, machine-checkable rulebook can certify every safe program an AI assistant might write, because properties like "never" and "eventually" cannot be settled by a finite record. The best a gatekeeper can offer is a checkable list that includes everything it must allow and nothing it must refuse.
3. The authors argue that constitutions for sovereign AI models should evolve as participatory documents, with provisions marked as settled, contested, or underdetermined. Agent-based simulations suggest the process holds up even when half of citizens are bad actors, assuming 75% of deliberators act in good faith.
4. SafeRoute-MAS decides, using only pre-integration information, whether to keep all specialist messages or drop one in multi-agent scientific reasoning. It gained 1.80 points on a cached set while cutting 12.76% of message tokens, but controlled integrations on gpt-oss-120b did not establish a selector advantage, so the authors propose a Keep-All fallback.
5. NTT announced a physics-based world model that checks candidate robot actions before execution to prevent failures like dropped objects. In a small-robot shelf-loading test, task success rose from 73.1% to 92.3%, and the first planned use is airport baggage loading.
6. This paper pairs online-adaptive fuzzy cognitive maps with LSTM networks for financial forecasting. On NASDAQ Composite data, the hybrid matched the LSTM baseline on directional accuracy and beat it by 7.6 points on return-sign accuracy, with its maps adapting to new regimes within 10 to 15 trading days.
7. Veritas is a harness that automatically generates verifiers for agent tasks, which usually must be written by hand. Its verifiers agree better with official verdicts than those of seven other harnesses on all four execution-scored benchmarks.
8. MSPR-GRE is a self-supervised reconstruction method for accelerated multi-echo gradient-echo MRI that trains without fully sampled targets. At 12x acceleration, fourth-echo magnitude error was 12.09% versus 21.55% for the SSDU baseline.
9. The paper separates the spatial structure inside images from the cross-modal positional effects that Multimodal Rotary Positional Embedding imposes on text tokens in vision-language models. It finds the cross-modal effects add little and can impair performance, while intra-visual encoding remains useful for fine-grained grounding.
10. The authors propose studying AI agents with epidemiologic cohort designs, treating agents as members of a cohort and their behaviors as exposures that may affect later outcomes. They illustrate the approach with longitudinal studies of curiosity and probing as a safety question and data checking as a reliability question.
11. MPIUR is a multimodal graph-based recommender that combines item descriptions, images, and user reviews in one framework. It achieved the best nDCG@10 and Recall@10 on two Amazon Reviews 2023 categories, including on sparse items with limited interaction histories.
12. F3-AGI v4.0 is a CPU-only neuro-symbolic architecture for ARC-AGI that makes no LLM calls. It solves 57 of 400 public evaluation tasks, at a mean latency of 1.62 seconds per task, and 38 of 350 strictly held-out tasks.
13. This study annotates 95 failed runs from six multi-agent systems and finds that faults usually enter early in a run. In 72% of failed runs, the trajectory already contained evidence against the error, yet in 88% of those cases no agent acted on it.
14. ExpertCrowd-SWE is a benchmark of 100 expert-written, multi-file engineering tasks across 87 open-source projects. Claude Opus 5.5 (67.3%) and GPT-6 Astra (63.7%) lead without a statistically significant gap, and no task is fully solved by every model.
15. Mercor announced a $750 million Series D at a $20B valuation, with a $3B run rate. The brief's summary line also mentions Google's general availability of "Gemini 4 Carbon," though the source gives no further detail.
16. R2C-Agent is a decision-making agent for HIV interventions on hidden contact networks, observing only the peer referral networks that are cheaper to collect at scale. It uses a conditional graph diffusion model to reconstruct likely contact networks and outperforms baselines in simulations on real-world HIV network data.
17. This Google blog post highlights the art and legacy of Dr. Esther Mahlangu. The extracted text is only a photo caption, so no further details are available here.
18. Large language model agents, connected to a GPU evaluator through the Model Context Protocol, proposed a 13-parameter exchange-correlation functional called O159. After self-consistent evaluation, it lowers the held-out MGCDB84 test error of ωB97M-V by 11.7%.
19. This paper studies measurement when AI labels must stand in for human readings of 121,070 US tax-filing grant-purpose texts. A two-phase human review policy cut the error of the estimated monetary total by 28% at a 100-text budget, and independent human readers shifted the estimate by $3.93B.
20. The Neural Control Plane observes a language model's execution, synthesizes temporary weight-space interventions, and keeps a versioned table of admitted policies. In a SmolLM-135M test, one projection halved token repetition, but no candidate produced new correct answers, so the admission gate kept the original policy.
21. W-ISA is a software instruction set and transactional runtime for applying weight-space interventions to pretrained language models. It restores saved parameters after each request, and on WikiText-2 the combined ablation raised perplexity by only 0.38%.
22. This paper presents a retrieval-grounded framework for medical multiple-choice questions that extracts evidence distinguishing confusable answer options. It shows consistent accuracy gains across several datasets and target LLMs, outperforming both raw retrieved chunks and general LLM-generated hints.
23. Armada started continuous production at Galleon Forge Two, a 90,000-square-foot facility inside a Johnson Controls plant in Dublin, its first European manufacturing operation. The first Dublin-built Leviathan modular data center was completed for Fossefall in Norway.
24. This study tests Claude Code, an agentic deployment of Claude Opus 4.6, as a grader of handwritten university exams. Its totals agreed closely with the expert's (r = 0.97), but the authors conclude it can assist on objective items yet is not ready to replace instructors on boundary cases or drawing questions.
25. LG-HRL is a language-grounded hierarchical reinforcement learning framework for coordinating UAV and UGV search-and-rescue teams from natural-language instructions. In a simulated disaster environment it reports 2.3 times faster learning and 20 to 30% higher victim rescue rates.

Summaries written with AI from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering
2. No finite rulebook can certify every safe program: an impossibility theorem for gatekeeping AI agents
3. Living constitutions: Participatory alignment for sovereign foundation models
4. SafeRoute-MAS: Coverage-Aware and Risk-Calibrated Message Routing for Multi-Agent Scientific Reasoning
5. NTT’s Physics-Based World Model Screens Robot Actions Before Execution
6. Online-Adaptive Fuzzy Cognitive Maps for Neuro-symbolic Financial Forecasting
7. Veritas: A Verifier Generation Harness for Agent Tasks
8. Multi-Stage Phase-Aware Self-Supervised Reconstruction for Multi-Echo Gradient-Echo MRI
9. Disentangling Multimodal Rotary Positional Embedding in Vision-Language Models
10. Do AI Agents need Preventive Care? An Epidemiologic Cohort Design for Agentic AI
11. A Technique to Model Product Information and User Reviews (MPIUR)
12. F3-AGI v4.0: Bioenergetically-Gated Active Inference, Metaplastic Preconditioning, and Residual Program Synthesis for Sample-Efficient Abstract Reasoning
13. Dependable Long-Horizon LLM Agents: A Trajectory-Centered Empirical Study of Where Errors Enter and Which Controls Reach Them
14. ExpertCrowd-SWE: Measuring Frontier Coding Agents on Expert-Authored Engineering Tasks in Unfamiliar Codebases
15. [AINews] Mercor $750m Series D at $20B valuation, $3B run rate
16. Towards Scalable Network-Based Interventions for HIV Treatment and Prevention via R2C-Agent
17. Experience the art and legacy of Dr. Esther Mahlangu
18. Agent-Guided Discovery of Exchange-Correlation Functionals with Physics-based Validation
19. Counting Under Disagreement: Dollar-Weighted Measurement, Selective Machine Labelling and Budgeted Human Review from Disagreeing AI Readers of Administrative Text
20. NCP: Verifier-Guided Optimization of Weight-Space Control Policies
21. W-ISA: Transactional Execution of Weight-Space Intervention Programs
22. Decisions by Discrimination: Differential Evidence for Medical Reasoning
23. Armada Starts Dublin Production of Sovereign AI Data Centers
24. Reliability of Claude Code as an Agentic AI Grading System for Handwritten University Exams
25. Cognitive Multi-Agent Search and Rescue: Language-Grounded Hierarchical Reinforcement Learning for Collaborative Robotics