AI News Brief Roundup: 2026-11-10 Morning

AI News Brief: Morning roundup, 2026-11-10

Summary

This batch is dominated by arXiv preprints, and the common thread is the gap between benchmark performance and dependable behavior. Several papers show safety and reliability failing in ways that are easy to miss: fine-tuning can erode refusals, poisoned training data can hide backdoors, and clinical summaries can quietly drop important details. Efficiency work on serving large models and a few essays questioning whether scale alone will produce intelligence round out the picture, suggesting that capability gains are real but verification and oversight are still catching up.

1. On 18,000 adversarial CT and MRI summary pairs, GPT-5-mini was the most faithful model, yet it still hallucinated in 28-30% of summaries under term substitution. Claude Sonnet 4.5 and MedGemma did worse, at roughly 35-36% and 41-42%.
2. SPECTRA restores refusal behavior in fine-tuned LLMs without retraining, by transferring the base model's refusal direction into the fine-tuned weights through a low-rank, activation-aware update. It needs no gradient optimization or domain-specific safety data and leaves false refusal rates nearly unchanged.
3. TripleGrader is an open-source web platform for grading knowledge-graph triples on a 0-10 scale, with inter-rater agreement metrics built in. When LLM-extracted triples were curated into two gold benchmarks, only 10-32% could be kept unchanged.
4. Testing transitivity across pairwise moral trade-offs, the authors found that most LLM choice sets produced weak or unstable orderings. Averaging responses, a common evaluation practice, created structure that matched no individual choice set.
5. Simon Willison compared building a SQLite extension that parses strings like -$1,234.56 in C and in Rust. The Rust builds were 298-321 KB against 14 KB for C, and a C parser using Clinger's fast path finally beat Rust on speed, with Claude Opus 5.5 writing the code.
6. FairRec is an item-level decoding framework that keeps LLM recommendations stable when a protected attribute changes, without access to model weights or full logits. Its soft variant lowered counterfactual disparity compared with zero-shot prompting, UP5 and FACTER, while keeping accuracy competitive.
7. AdaOffload is a multi-GPU framework for mixture-of-experts inference that splits expert execution between GPUs and host CPUs, addressing models too large for GPU memory. Across five hardware platforms and three open-weight models of 284B to 753B parameters, it reported 1.9x to 15.5x throughput gains over existing offloading systems.
8. ScreenProof asks whether a model can flag UI faults such as clipped labels or broken line breaks without a reference design. Frontier judges falsely flagged 18-62% of clean real screenshots, while the authors' small ProofJudge, a fine-tuned Qwen3.5-9B, achieved the highest F1 score of the judges tested.
9. CoRA-Bench measures whether context selectors keep the evidence needed to answer a question when inputs must fit a token budget. BM25, dense retrieval, reranking and compression methods were compared, and their rankings shifted across evidence types and between evidence recall and final answer accuracy.
10. Mihail Eric, from the engineering team at Monaco, discusses how Stanford's CS146S course is taught when coding agents can finish last year's assignments in one prompt. The conversation looks at what AI-native engineering education should look like.
11. ProofDiff builds source-grounded argument maps for two informal mathematical proofs and proposes correspondences between them, showing whether the documents follow the same reasoning route. It is designed to support human review rather than certify correctness, and was demonstrated on four pairs of AI-generated proofs.
12. Mihail Eric describes a second run of Stanford's CS146S in which open-source maintainers joined the course, so students had to learn an unfamiliar repository and propose a change. The course is redesigned around the fact that coding agents can now complete traditional homework quickly.
13. In "Is Recursive Self-Improvement Safe?", the author argues that clear frameworks for predicting the outcomes of AI systems that improve their successors do not yet exist, even as frontier abilities make the idea seem plausible. The piece adds a quick game-theoretic analysis of whether to start such a process when a rival is making the same choice.
14. PISA is a prompt-adaptive, cost-aware router that picks one model per request, and PISA-E extends it to ensembles whose composition adjusts to each query. PISA-E outperformed ensemble baselines on task performance and, at lower-cost settings, reduced inference cost by roughly 36% and 25% in in-distribution and out-of-distribution tests.
15. EvoDiff is a generative active learning framework for ultrasound that uses diffusion-generated probes to retrieve only the most informative real images for annotation. Across 15 datasets and 16,696 images it approached fully supervised performance with fewer labels, and a six-radiologist reader study estimated a 24-84% cut in annotation workload.
16. The authors argue that independent verification organizations, a popular bipartisan idea for AI evaluation, should be designed for everyday system- and application-level risks, not just catastrophic frontier risks. They draw on finance, education, public health and science to outline who these organizations would serve, what access they need and how they might be funded.
17. AutoEMatic adjusts the precision of weights, activations and the KV cache for each layer at runtime, using next-token entropy and the size of each layer's hidden-state update. It reports accuracy gains of up to 21.7% over NVFP4 at 4.9 bits per weight, and 1.6-2.3x lower latency on a custom accelerator at matched accuracy.
18. AVCORP is a corpus of 5,000 annotated utterances from 250 games of the deception game Avalon, used to train a deception detector on Qwen3-8B. On held-out games the trained model outperformed its untrained baseline and frontier models including Gemini 3.1 Pro and DeepSeek-V4-Pro, matching Claude Opus 5.5.
19. This config-verified comparison covers ten open mixture-of-experts models, including DeepSeek, Qwen, Mixtral, Llama 4 and gpt-oss. It lays out total and active parameters, expert counts, top-k values, shared experts, gating functions and load-balancing methods, making it a practical reference for how each model routes tokens.
20. WorldMind lets a vision-language model choose spatial transformations that a world model executes, feeding latent features back into reasoning rather than decoding imagined states into pixels. It reports an average 16.0% gain over its base VLM across VSI-Bench, MMSI-Bench and MindCube-Tiny, and 8.3% over the strongest RGB-based world simulation.
21. In a Nature commentary, consciousness researchers Anil Seth, Ned Block and Giulio Tononi argue that today's AI systems are not conscious and will not become so on silicon hardware. They call the assumption that consciousness depends on computation alone "fatally flawed" and warn against granting rights to AI systems.
22. A juried in-person AI art exhibition with 209 artists, lawyers and computer scientists found that nearly half of evaluators would buy work, with prices tracking perceived creativity and visual appeal. Three-quarters of creators would sell at prices tied to effort rather than artistic value, and both groups saw prompting users and AI developers as deserving rewards, leaving training-data artists out.
23. Google's Elephant-Goldfish plugin makes coding agents draft a specification first, then pass it through five review stages that each run without the surrounding context before any code is written. It is aimed at spec-first coding workflows.
24. The "homeopathic learning" paper shows that poisoned fine-tuning data with little effect on its own can create a hidden backdoor once clean, unrelated data is added, because the trigger becomes the attributed cause. The authors demonstrate this across backdoor, reward-hacking and mislabeled-MNIST settings, suggesting that auditing or filtering training data in isolation can miss the risk.
25. Turing Award winner Richard Sutton argues that raw compute alone is not enough for intelligence and that fundamentally new learning methods are needed. Former OpenAI researcher Tworek adds that blind transformer scaling "has reward hacked brains of a whole generation of scientists," and a growing set of startups is betting on that gap.

Summaries written with AI from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. Adversarial Evaluation of Large Language Models for Patient-Friendly Radiology Report Summarization
2. Restoring General and Improving Domain-Specific Safety in Fine-Tuned LLMs Without Retraining
3. TripleGrader: A Web-Based Platform Supporting Collaborative Triple Annotation with Quality Grading and Inter-Rater Agreement Assessment
4. Do large language models rank moral values? A transitivity test of moral competence
5. parse_number() as a SQLite extension: C vs Rust
6. Conformal Fair Decoding for Large Language Model Recommenders
7. Multi-GPU MoE Inference with Adaptive Expert Offloading
8. ScreenProof: Where Is This Screen Wrong? Finding UI Faults Without a Reference
9. CoRA-Bench: Evaluating Evidence Preservation by Context Selectors Under Token Budgets
10. How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco
11. ProofDiff: Auditable Argument-Level Comparison of AI-Generated Mathematical Proofs
12. [One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco
13. Is Recursive Self-Improvement Safe?
14. Ensemble Routing for Prompt-Adaptive Cost-Aware LLM Inference
15. Generative Active Learning for Label-Efficient Pan-Organ Ultrasound Diagnosis
16. Key Institutional Considerations for Designing Effective Independent Verification Organizations
17. AutoEMatic: Runtime-Adaptive Mixed-Precision LLM Inference via Automatic Exponent-Mantissa Refinement
18. AVCORP: Identifying and Resisting Deception via Communication Theory and Theory of Mind
19. Which Routing Does Each MoE Model Use? DeepSeek, Qwen, Mixtral, Llama 4, gpt-oss Compared
20. WorldMind: Test-Time Scaling for Visual Spatial Reasoning through Latent World Simulation
21. The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate
22. What Is the Value of AI-Involved Art? Evidence From a Juried In-Person Exhibition
23. Google AI Releases Elephant-Goldfish Model Plugin for Spec-First Coding
24. Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect
25. The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts