1. Podcasts
  2. Technology
  3. Best AI papers explained
Best AI papers explained cover art

530 episodes

Best AI papers explained, page 15

by Enoch H. Kang · English

Cut through the noise. We curate and break down the most important AI papers so you don’t have to.

All episodes, page 15

Reusing pre-training data at test time is a compute multiplier

The episode discusses a research result that reusing the model’s pre-training data at test time via retrieval-augmented generation (RAG) acts like a compute multiplier, quantifying inefficiency in…

10 Nov 2025 · 16m · 8 chapters

Scaling Agent Learning via Experience Synthesis

The episode explains why large-scale training of LLM-based autonomous agents is blocked by “experience bottlenecks” in reinforcement learning, and how DreamGym addresses this by synthesizing…

9 Nov 2025 · 17m · 10 chapters

Continuous Autoregressive Language Models

Continuous Autoregressive Language Models (CALM) aim to reduce LLM inefficiency from token-by-token next-token prediction by predicting the next vector for a chunk of K tokens (“semantic bandwidth”),…

8 Nov 2025 · 16m · 8 chapters

Toward a Theory of Agents as Tool-Use Decision-Makers

The episode explains a research framework for “agents as tool-use decision-makers,” aiming for optimal, efficient behavior by aligning when an agent uses internal reasoning tools versus external…

7 Nov 2025 · 20m · 9 chapters

Nested Learning: The Illusion of Deep Learning Architectures

The episode argues today’s large language models are “static” after pretraining because they lack rapid online memory consolidation, leading to an “anterograde amnesia” effect.

5 Nov 2025 · 13m · 6 chapters

GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding

GST-UNet, a neural framework for spatiotemporal causal inference with time-varying confounding, combining UNet-style spatial modeling with iterative G-computation to estimate counterfactual effects…

5 Nov 2025 · 18m · 8 chapters

Beyond a million tokens: benchmarking and enhancing long-term memory in llms

Long-term memory in LLMs beyond million-token context windows; introduces Beam benchmark and Light memory architecture to reduce contextual forgetfulness and improve long-range coherence.

4 Nov 2025 · 15m · 10 chapters

Agentic Economic Modeling

Agentic economic modeling (AEM) uses LLMs as “economic agents” to generate counterfactual decision data, then corrects systematic LLM bias with a small amount of real human data, enabling econometric…

3 Nov 2025 · 14m · 7 chapters

Emergent Introspective Awareness in Large Language Models

Whether large language models’ self-reports about “thoughts/feelings” are genuine introspection or confabulation, and how Anthropic researchers test “emergent introspective awareness” using concept…

3 Nov 2025 · 16m · 7 chapters

Can Large reasoning models self-train?

Self-rewarded training (SRT) for large reasoning models—letting models “grade their own homework” using self-consistency/majority voting to create pseudo-rewards, aiming to reduce reliance on…

1 Nov 2025 · 12m · 4 chapters

ALITA-G: Self-Evolving Generative Agent for Agent Generation

Alita-G, a self-evolving generative agent framework that turns a general-purpose agent (“masquerade”) into a reusable domain expert by generating, abstracting, and curating specialized toolkits (“MCP…

1 Nov 2025 · 16m · 10 chapters

Self-improving LLM agents at test-time

Test-time self-improvement (TTSI) for LLM agents—learning during inference only for uncertain cases, using a three-step loop (H self-awareness, G self-data augmentation, T parameter-efficient…

30 Oct 2025 · 19m · 8 chapters

Offline RL by Reward-Weighted Fine-Tuning for Conversation Optimization

Offline reinforcement learning for LLM conversation optimization via reward-weighted fine-tuning.

30 Oct 2025 · 15m · 9 chapters

Language models are injective and hence invertible

The episode argues that standard decoder-only transformers are structurally lossless because the mapping from prompt tokens to internal hidden states is (almost surely) injective, so hidden states…

30 Oct 2025 · 12m · 5 chapters

ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory

ReasoningBank, a Google Cloud AI research memory framework for LLM agents that prevents “forgetfulness” by storing structured, transferable reasoning strategies distilled from both successes and…

29 Oct 2025 · 15m · 11 chapters

RLAD: Training LLMs to Discover Abstractions

RLAD (reinforcement learning with abstraction discovery) trains LLMs to solve hard, unseen reasoning problems by first generating concise “reasoning abstractions” (high-level strategies/guardrails)…

29 Oct 2025 · 16m · 15 chapters

How to Train Your Advisor: Steering Black-Box LLMs with ADVISOR MODELS

Advisor models for steering “black-box” LLMs via a trainable small model that generates per-request natural-language steering instructions, optimized with reinforcement learning while the large model…

29 Oct 2025 · 13m · 7 chapters

Self-improving LLM agents at Test-Time

The episode discusses test-time self-improvement (TTSI) for LLM agents, aiming to reduce information overload and agent training cost by adapting only when the model is uncertain.

27 Oct 2025 · 23m · 11 chapters

KL-Regularized Reinforcement Learning is designed to Mode Collapse

KL-regularized reinforcement learning (RL) fine-tuning for LLMs can mathematically force mode collapse, reducing output diversity.

27 Oct 2025 · 16m · 8 chapters

How do LLMs use their depth?

How LLMs use their depth during inference, arguing for a “guess then refine” workflow plus “complexity-aware depth use” (more layers for harder parts).

27 Oct 2025 · 12m · 7 chapters

Thought Communication in Multiagent Collaboration

Thought communication (ThoughtCom/TCOM) for multiagent collaboration, enabling LLM agents to share latent “intent drivers” directly instead of exchanging language tokens, avoiding language…

27 Oct 2025 · 17m · 10 chapters

Reasoning with Sampling: Base Models Outperform RL

The episode argues that “power sampling” can make base LLMs outperform or match RL post-training methods (e.g., GRPO) for reasoning tasks using only inference-time sampling—no new training, data,…

26 Oct 2025 · 16m · 10 chapters

Continual Learning via Sparse Memory Finetuning

Continual learning for LLMs using sparse memory fine-tuning (SMF) to prevent catastrophic forgetting.

26 Oct 2025 · 14m · 5 chapters

Direct Preference Optimization with Unobserved Preference Heterogeneity: The Necessity of Ternary Preferences

Pluralistic AI alignment using Direct Preference Optimization when user preferences vary in unobserved ways; argues standard binary preference data (A vs B) is mathematically insufficient and…

24 Oct 2025 · 12m · 9 chapters