

531 episodes
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.

PLUS (pluralistic alignment via summarization) improves RLHF personalization for diverse users by replacing a single-average Bradley-Terry-Luce reward model with user-specific, text-based preference…
4 Oct 2025 · 16m · 8 chapters
How RLHF preference learning hides “hidden context” from annotators and, via its aggregation math, implicitly implements a social choice rule (Borda count), leading to strategic manipulability and…
3 Oct 2025 · 16m · 8 chapters
“Age of AI Agency” and the LIME/LiMI research (“Less is More for Intelligent Agency”) arguing that autonomous AI agents can be trained with far fewer examples by using high-quality “trajectories”…
1 Oct 2025 · 14m · 5 chapters
LoRA (low-rank adaptation) for adapting large language models efficiently, focusing on when it matches full fine-tuning (“low regret”), plus operational benefits, compute/FLOPs savings, and behavior…
1 Oct 2025 · 22m · 9 chapters
Actor-Critic Without Actor (ACA) for reinforcement learning replaces the usual actor network with critic-guided denoising, aiming to remove policy lag and cut model size while retaining multimodal…
29 Sep 2025 · 16m · 11 chapters
Whether reinforcement learning (RL) can make large language models (LLMs) acquire genuinely new programming algorithms, and whether that skill transfers beyond the training distribution.
29 Sep 2025 · 16m · 6 chapters
A linear transformer trained on masked matrix block completion implicitly discovers a unified numerical solver, EGLE (Emergent Algorithm for Global Low-Rank Estimation), a two-line update rule…
29 Sep 2025 · 14m · 7 chapters
Regularizing extrapolation in causal inference when source and target groups differ (positivity violations).
27 Sep 2025 · 15m · 10 chapters
DoubleGen, a framework for generating unbiased counterfactuals (“what if” outcomes) from biased observational data by debiasing confounding and avoiding misspecification.
27 Sep 2025 · 13m · 6 chapters
Effective reasoning in large reasoning models (LRMs) is driven by structure and failure management, not by longer chain-of-thought (CoT) or more “review” tokens.
27 Sep 2025 · 17m · 10 chapters
Compute as Teacher (CAT) turns a model’s inference-time exploration into reference-free supervision for training specialized skills, reducing the “supervision gap” when gold labels are scarce or…
27 Sep 2025 · 16m · 8 chapters
Explains in-context learning (ICL) as implicit optimization during inference, not “magic” or explicit weight updates.
24 Sep 2025 · 14m · 7 chapters
Tests whether reinforcement learning with verifiable rewards (RLVR) truly increases LLM reasoning capacity beyond the base model, or only improves sampling efficiency within existing capabilities.
24 Sep 2025 · 13m · 7 chapters
Open problems in mechanistic interpretability (MI): how to reverse-engineer neural networks’ internal computations to enable trust, control, and scientific discovery despite “black box” behavior.
21 Sep 2025 · 19m · 12 chapters
The episode explains a study on “thought anchors,” sentence-level reasoning steps that disproportionately determine how an LLM reaches its answer, making chain-of-thought more interpretable and…
21 Sep 2025 · 12m · 5 chapters
“Thought anchors” for LLM chain-of-thought—identifying which reasoning sentences most causally determine final answers in hard math, using intermediate-level abstraction rather than token-level…
21 Sep 2025 · 16m · 9 chapters
Catastrophic forgetting in continual learning, and a MIT paper’s claim that “RL’s razor” explains why online reinforcement learning (RL) forgets less than supervised fine-tuning (SFT).
7 Sep 2025 · 25m · 12 chapters
Why large language models hallucinate (produce plausible but false statements) and why hallucinations persist due to training objectives and benchmark evaluation incentives; proposes changing…
6 Sep 2025 · 18m · 8 chapters
ALFA (Alignment Via Fine-Grained Attributes) teaches LLMs to ask better follow-up questions for clinical reasoning by decomposing “good questions” into attributes, synthesizing targeted training…
6 Sep 2025 · 16m · 12 chapters
The episode discusses the paper “Sample Efficient Preference Alignment in LLMs via Active Exploration,” arguing that aligning LLMs to be helpful/harmless is costly because it requires many human…
6 Sep 2025 · 15m · 5 chapters
Using AI to estimate consumer demand price elasticity from product details, improving on traditional econometric methods that understate price effects due to confounding factors.
4 Sep 2025 · 14m · 5 chapters
Memento proposes “fine-tuning LLM agents without fine-tuning LLMs” by enabling continuous, real-time learning for autonomous LLM agents using an external memory (case bank) rather than retraining the…
1 Sep 2025 · 19m · 9 chapters
Theoretical and empirical limits of embedding-based (single-vector) retrieval models, showing they cannot represent all possible “top-K relevant document sets” for a query unless the embedding…
31 Aug 2025 · 17m · 12 chapters
A Google Research approach for predicting performance of large, complex industrial systems using text-to-text regression with regression language models (RLMs), avoiding lossy tabular feature…
30 Aug 2025 · 16m · 8 chapters