Deep Learning Theory Maturing
Key Questions
What theoretical frameworks are advancing understanding of LLMs?
Recent work connects statistical physics, spontaneous symmetry breaking, random matrix theory, and control theory to explain LLM behaviors like grokking and double descent. Papers on Hamilton-Jacobi theory and geometric latent reasoning provide quantitative tools for generalization and robustness.
How are self-improving AI agents evolving according to new research?
New methods like Self-Revising Science Agents, SIA, and HarnessForge enable agents to detect knowledge gaps, co-evolve policies with harnesses, and achieve gains such as 56.6% on LawBench. These approaches move from routine search toward principled discovery using concepts like category theory.
Why is the ICML position paper on AI deception important?
The paper by Mitchell et al. distinguishes genuine deception from role-play in AI systems, calling for more rigorous experimental design in alignment research. It has been highlighted by researchers as critical for understanding misalignment risks.
Stat physics phase trans to LLMs/Hopfield; spontaneous symmetry breaking (Bronstein); non-eq DMFT/double descent; RMT/grokking; EBMs (LeCun); Rigorous Theory of LLMs (Brock); control theory; Sara Hooker post-scaling; unifying physics/neuroscience/AI. New: Hamilton-Jacobi Theory of Deep Learning (exact PDE correspondence, quantitative generalization/robustness/influence functions); Geometric Latent Reasoning (GLR) induces shorter generations; 'Why Larger LLMs Learn Rare and Complex Tasks' (reduced interference); Local Perturbation Theory for Multi-Domain RL (second-order effects); Influence-Guided Symbolic Regression (LLM+MCTS); Shay Moran talk on geometry of ranking; Latent Prediction beats token-level training (exponential sample efficiency); Omar Sar follow-up (evolver plateaus, solver inverted-U); Jing Huang paper questioning larger models; MUX method (latent continuous reasoning); Meta-Agent Challenge (frontier models struggle to self-improve); Noam Razin talk on proxy reward functions; AI solves Grothendieck constant; Meta-Cognitive Memory Policy Optimization (97.1% at 1.75M tokens); Shadow Price of Reasoning (CLEAR 3x accuracy). Also MARS automates AI research; STRIDE training data attribution via activation space. New: Self-Revising Science Agents via Category Theory (copresheaves to distinguish routine search from true discovery); @blader announces breakthrough in self-evolving AI scientists moving from search to principled discovery. SePO introduces self-referential evolutionary system prompt optimization, showing consistent gains and pre-training generalization. SIA unifies harness and weight updates in a single self-improving loop (56.6% LawBench gain, 502% denoising). OpenSkill bootstraps skills and verification from scratch for open-world self-evolution without target-task supervision. New: HarnessForge formalizes harness-policy co-evolution for LLM agents (12% gain). KnowSelf introduces self-knowing agents that detect knowledge gaps and strategically acquire skills. New: Experience Makes Skillful (SkeMex) for medical agent self-evolution with structured skill memory; SEE elicits latent judge calibration with minimal data (160 examples); PBSD addresses long-horizon credit assignment via Bayesian self-distillation. Also Math Theory of Deep Representation Learning (unifying memory and world models). Cosine Misleads paper challenges latent visual reasoning assumptions in VLMs. New agent self-improvement papers: Role-Agent (dual-role evolution), RHO (retrospective harness optimization, 59%→78% on SWE-Bench Pro), SearchSwarm (delegation intelligence, SOTA on BrowseComp), EEVEE (test-time prompt learning for heterogeneous streams). Also UC San Diego Professor Daniel Kane awarded Gödel Prize for robust high-dimensional statistics. New: Deficient executive control in transformer attention paper identifies fundamental failure mode. EvoTrainer co-evolves LLM policies and training harnesses for autonomous agentic RL. New: Daniel Barzilai (Weizmann) talk on model collapse and SGD lower bounds—model collapse positive under regularity, negative without structure; SGD lower bounds via direction-finding difficulty. Edward Lockhart (DeepMind) talk on formal mathematics as solution to reward hacking. New: HarnessBridge introduces learnable bidirectional controller for LLM agents, extending harness-policy co-evolution. New: OPD geometry paper reveals sparse, FFN-concentrated updates with spectral signatures. New: Luca Baggi talk on mechanistic interpretability synthesizes recent work on circuit tracing and sparse features. New: ICML position paper on distinguishing genuine deception from role-play in AI (Mitchell et al.) forces more rigorous experimental design in alignment research.