AI Briefing: November 4, 2026
Today's briefing covers breakthroughs in weight-space programming, energy-efficient inference, and agent safety, alongside industry moves from OpenAI, Anthropic, and major cloud providers.
Models and Research
LLMs can read & write neural network weights: 'Weight Space Programming' cuts Autoresearch out of the RSI loop
LLMs can directly read and edit neural network weights, enabling bug fixes and weight generation without traditional training, marking a step toward recursive self-improvement.
SIMPLE Law: A Closed-Form Analytical Model for LLM Inference Energy Cost
SIMPLE law predicts LLM inference energy from architecture with high accuracy, revealing quadratic prefill growth and 13x energy increase for agentic serving.
SpecOPD: Branch-Parallel On-Policy Distillation for Speculative Decoding
SpecOPD enables branch-parallel on-policy distillation for speculative decoding, boosting acceptance length by 41% and generation speedup up to 3.56x.
Predictive Safety Guardrails Do Not Always Make LLM Agents Safer
Predictive safety guardrails don't always improve LLM agent safety, with stronger prediction involvement reducing task completion and increasing costs.
How “reverse game theory” could keep AI agents in check
Mechanism design offers a new alignment approach, using incentives to make rule-following rational for AI agents, though challenges like collusion remain.
Same Outputs, Different Scores: Auditing the Unchecked Graders of Multimodal LLM Benchmarks
Audit of multimodal LLM benchmark scorers shows accuracy swings up to 22 points, urging standardized reporting of non-compliance and multiple scoring rules.
SQLoom: Set-Contrastive Learning for Text-to-SQL
SQLoom improves text-to-SQL with set-contrastive learning, boosting Qwen3-1.7B accuracy from 62.9% to 88.7% on Spider dev.
Auto-Research Agent Optimizes MSA for Protein Complex Prediction
Auto-research agent optimizes MSAs for protein complex prediction, improving interface DockQ while using 60% fewer sequences.
Global pooling beats per-ship and adaptation models for ship fuel consumption prediction
Global pooling models beat per-ship and adaptation approaches for ship fuel prediction, achieving 8.92% MAPE and enabling zero-shot prediction.
LANCE: Bridging Low-Level Image Fusion and High-Level Semantic Understanding
LANCE bridges image fusion and semantic understanding via structured language, improving downstream detection and F1 scores on remote sensing tasks.
JABER: Sequential Quality Control for Human-AI Data Annotation under Shared Bias
JABER reduces QC loss by up to 10% in human-AI annotation under shared bias, achieving zero labeling errors on FEVER.
Integrating Domain Adaptation and Retrieval-Augmented Generation for Saline-Alkali Land Reclamation Planning
Wenyao combines domain adaptation and RAG for saline-alkali land planning, achieving 95.83% accuracy on domain-specific questions.
KAN vs. MLP Benchmark on Clinical Tabular Data Shows Trade-offs in Precision vs. Efficiency
KAN variants improve minority-class detection but require more parameters and inference time; MLPs remain efficiency Pareto frontier.
Value-Conditioned Proposal Acquisition for Test-Time Reasoning
ValueFormation improves test-time reasoning by acquiring proposals from specialists, boosting MATH-500 accuracy by over 15 points.
Math Foundation Formalization Agent: A General-Purpose Approach to Formalizing Mathematical Foundations
MFF-Agent automates formalization of mathematical foundations with LLMs and Rocq, demonstrating feasibility across diverse domains.
CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents
CARE enables recursive self-improvement in production customer-service agents, learning from failures with human approval.
Creating Hybrid Interactive Narrative Systems with LLM: From Formal Description to Evaluation
LLM-based narrative platform compares three Experience Managers, showing no single manager dominates across agency, controllability, and robustness.
OmniOPD: Stable On-Policy Distillation for Joint Audio-Video Generation
OmniOPD stabilizes joint audio-video generation with modality-wise on-policy distillation, preserving audio quality while enhancing video.