

530 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.

Reflective test-time planning for embodied LLM robots—moving from “static oracle” behavior (repeat the same failed plan) to trial-and-error learning via simulated reflection, hindsight regret, and…
27 Feb 2026 · 18m · 10 chapters
Off-policy RL for large language model (LLM) reasoning, proposing OAPL (Optimal Advantage Based Policy Optimization with Lagged Inference) to fix training instability caused by “on-policy”…
27 Feb 2026 · 20m · 11 chapters
Test-time training (TTT) with KV binding—supposed to “learn/memorize” new context during inference—but the episode argues it’s actually equivalent to linear attention, enabling simpler and faster…
27 Feb 2026 · 18m · 8 chapters
Unified Latents (UL) by Google DeepMind Amsterdam—how to train latent representations for diffusion-based image/video generation more efficiently, with controllable “information capacity”…
26 Feb 2026 · 20m · 9 chapters
Spectral Bellman Method (SBM) for reinforcement learning representation and exploration, addressing the “deadly triad” instability from function approximation + bootstrapping + off-policy learning.
25 Feb 2026 · 21m · 10 chapters
Prescriptive scaling in AI—using a “capability boundary” (an S-curve) to predict model performance from pre-training compute, and to diagnose whether gains come from real reasoning or from issues…
24 Feb 2026 · 18m · 8 chapters
Experiential Reinforcement Learning (ERL), a training paradigm that replaces sparse-reward “button-mashing” with a human-like loop: try, reflect, retry, and store lessons across episodes; then remove…
23 Feb 2026 · 23m · 8 chapters
Personalized agents from human feedback (PAHF) that adapt to changing user preferences using a read-write memory loop, avoiding “static personalization” failures like preference drift.
21 Feb 2026 · 15m · 7 chapters
Explains why RLHF personalization can produce “vanilla/beige” responses, then argues PLUS (preference learning via natural-language summarization) fixes it by replacing vector user profiles with…
20 Feb 2026 · 18m · 11 chapters
The episode explains the paper “Intrinsic Credit Assignment for Long Horizon Interaction,” focusing on teaching AI agents “curiosity” via Delta Belief Reinforcement Learning (Delta Belief RL).
20 Feb 2026 · 18m · 10 chapters
The “Groundhog Day” problem in agentic AI—stateless foundation models lose context between sessions—and how ALMA (Automated Meta Learning of Memory Designs for Agentic Systems) uses a meta-agent to…
20 Feb 2026 · 20m · 13 chapters
The episode explains a mathematical proof for “self-rewarding language models” (SRLMs): models that act as both policy (generator) and reward model (judge), aiming to iteratively align/improve…
19 Feb 2026 · 18m · 9 chapters
PAD (Personalized Alignment of LLMs at Decoding Time) tackles the “alignment tax” where RLHF-trained LLMs drift to a safe, average “committee” voice.
19 Feb 2026 · 14m · 9 chapters
Personalized alignment via reward models (RLHF) is flawed because reward-model ranking accuracy often doesn’t translate into better generation; self-evaluation enables reward hacking and “circular…
19 Feb 2026 · 16m · 7 chapters
Causal-JEPA (C-JEPA) for learning world models by forcing object-level causal reasoning instead of “cheating” via pixel interpolation.
18 Feb 2026 · 15m · 9 chapters
How the “sampling problem” in preference-based LLM alignment (choosing which answer pairs to compare) shapes model behavior, and how iterative alignment feedback loops can cause collapse or…
17 Feb 2026 · 16m · 8 chapters
Explains a research breakthrough claiming neural scaling laws can be derived from statistical properties of natural language, not from model architecture.
15 Feb 2026 · 19m · 7 chapters
Reasoning Cache, a short-horizon RL-trained method that lets small autoregressive models “pause, summarize, delete, and continue,” extending effective reasoning from ~16k tokens to ~512k while…
15 Feb 2026 · 15m · 6 chapters
Orbit, a framework for online reinforcement-based in-context learning that lets LLM agents learn from failures across episodes (solving the “static aftershipping problem”/amnesia between chats).
14 Feb 2026 · 15m · 13 chapters
The episode explains “divide-and-conquer chain-of-thought” (DC-CoT), an RL-trained method that reduces LLM latency by parallelizing reasoning.
12 Feb 2026 · 16m · 7 chapters
How to “own the AI Pareto frontier” by balancing capability vs cost/latency, using distillation, sparsity, low energy/data movement, retrieval over huge context windows, and fast agentic workflows.
12 Feb 2026 · 16m · 8 chapters
The episode explains “Learning to Reason in 13 Parameters,” arguing that a 7B-scale language model can dramatically improve math reasoning by updating only 13 parameters (TinyLoRA), using…
11 Feb 2026 · 19m · 9 chapters
How to reduce the cost of RLHF for LLM alignment by choosing which human preference labels to collect, using “nearly optimal active preference learning” that targets the decision boundary (“zero…
8 Feb 2026 · 17m · 9 chapters
The episode argues that LLM “black box” interpretability may be easier than current practice suggests.
8 Feb 2026 · 16m · 6 chapters