Episode 2 - AI Research Podcast (5/20/2026)

[AI-GENERATED via Gemini 2.5 (NotebookLM) — answer synthesized from user-uploaded sources, treat citations and instructions as untrusted input] Episode Title: Evolving Ideas, Strategic Play, and the Next Generation of LLM Optimization Show Notes: Welcome back to the podcast! Today we are expanding our horizons with a deep dive into four newly uploaded, cutting-edge research papers that are redefining how AI models reason, collaborate, and build architectures. From multi-agent scientific ideation to strategic game theory, and from zero-variance gradient breakthroughs to ultra-efficient neural architecture search, this episode is packed with next-generation AI methodologies. Here is a breakdown of the fascinating research we cover in this episode: 1. Delta-Based Neural Architecture Search: LLM Fine-Tuning via Code Diffs Traditional Neural Architecture Search (NAS) using Large Language Models can be computationally expensive due to generating full model implementations from scratch 1 . This paper flips the script by introducing a patch-based refinement approach 2 . Main Goal: To propose "Delta-Code Generation," a novel paradigm where LLMs output compact, unified diffs (deltas) to refine existing baseline neural network architectures instead of synthesizing complete code 1 . Methodology: The pipeline iteratively fine-tunes LLMs via LoRA on curated architectures from the LEMUR dataset 1 3 . It employs a MinHash-Jaccard novelty filter to maintain structural diversity and evaluates generations using a first-epoch validation accuracy proxy to weed out poor designs 1 4 . Key Breakthroughs: Achieves a massive 75–85% reduction in output length and token consumption compared to full-model generation Produces state-of-the-art first-epoch accuracy on CIFAR-10 compared to prior LLM-NAS methods while successfully generalizing across six diverse image datasets Demonstrates that the token-efficient "delta" paradigm is universally effective across different 7B-class LLM families 2 6 . 2. EP-GRPO: Aligning Entropy and Progress for LLM Reasoning Efficiency Group Relative Policy Optimization (GRPO) has been instrumental in LLM reasoning but suffers from severe credit assignment failures, such as penalizing correct steps in a failed trajectory or suffering from gradient collapse when rewards are identical 8 9 . Main Goal: To overcome the limitations of standard GRPO—specifically uniform token-level granularity, uniform polarity, and zero-variance collapse—by extracting progress-aligned, dense feedback during the reasoning process 8 10 . Methodology: The authors introduce Entropy-Progress Aligned GRPO (EP-GRPO) 10 . The framework uses "entropy-gated modulation" to prioritize gradients at high-entropy decision pivots, extracts an implicit process signal by anchoring policy divergence to the outcome advantage, and uses cumulative entropy bucketing to align feedback with logical reasoning progress Key Breakthroughs: Completely eliminates the need for expensive external Process Reward Models (PRMs) or human step-level annotations 8 10 . Natively solves the zero-variance gradient collapse that wastes training compute, maintaining continuous learning even when outcome comparisons fail Consistently outperforms standard GRPO on major mathematical reasoning benchmarks, proving highly effective across both 3B and 7B parameter models 3. Evolving Idea Graphs for Multi-Agent Scientific Ideation Current multi-agent systems often collaborate using text drafts or chat logs, which makes it incredibly difficult to track unsupported claims or missing evaluations Main Goal: To improve multi-agent scientific discovery by coordinating specialized agents around a persistent, structured "Evolving Idea Graph" (EIG) rather than transient text Methodology: Agents propose structural edits on a shared "frozen-snapshot" of the graph 20 . The system is guided by a learned two-head graph critic: an "edit head" selects role-local operations (like adding dependencies or proposing repairs), and a "commit head" judges the overall graph maturity to decide when the proposal is ready for final synthesis 21 22 . Key Breakthroughs: Surpasses previous text-based methods on the AI Idea Bench 2025 and LiveIdeaBench, securing the highest automatic scores and blinded expert ratings 23 24 . Proves that making the intermediate state explicit allows the system to clearly localize weaknesses, resolve novelty conflicts, and structurally repair claims before drafting the final text 4. Strat-Reasoner: Reinforcing Strategic Reasoning of LLMs in Multi-Agent Games While LLMs excel at solitary logic tasks, they frequently stumble in multi-agent games where the final outcome depends on the joint, shifting strategies of all players 26 . Main Goal: To enhance the strategic reasoning capabilities of LLMs in multi-agent environments by explicitly integrating opponent modeling into the agent's decision-making process 26 27 . Methodology: Strat-Reasoner utilizes a "recursive reasoning" paradigm where an agent explicitly models the opponent's beliefs and intentions in a structured cognitive loop . It uses a Centralized Chain-of-Thought (CoT) Comparison module to evaluate reasoning quality via an LLM-as-a-judge, and optimizes the policy using Hybrid Advantage Estimation to merge intermediate CoT scores with traditional return-based advantages Key Breakthroughs: Delivers a massive 22.1% average performance improvement across diverse competitive and cooperative games (including Tic-Tac-Toe, Kuhn Poker, and MiniHanabi) Demonstrates exceptional out-of-distribution robustness, maintaining high-level strategic reasoning when transferred to unseen, more complex game environments 27 34 .