

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.

The episode analyzes why RLHF (reinforcement learning from human feedback) and DPO (direct preference optimization) can have different performance, focusing on “performance gaps” under model…
19 Jan 2026 · 14m · 9 chapters
The “reward engineering” failure mode in reinforcement learning—robots exploit poorly specified reward functions (e.g., dropping screws repeatedly) and multi-agent training breaks due to credit…
18 Jan 2026 · 18m · 12 chapters
Process Reward Learning (PRL) trains LLMs to improve reasoning by giving dense, step-by-step rewards instead of only rewarding final answers.
18 Jan 2026 · 16m · 10 chapters
The episode explains how “on-policy preference learning” can improve large language models faster than offline DPO by using self-generated, scored data.
17 Jan 2026 · 15m · 11 chapters
Stagewise reinforcement learning; why agents improve in sudden “jumps” due to a tradeoff between regret (performance) and internal complexity (geometry of the regret landscape), extending singular…
16 Jan 2026 · 13m · 3 chapters
Proposes “conditional memory” to complement Mixture-of-Experts (MoE).
16 Jan 2026 · 14m · 6 chapters
Learning latent action world models (LAMs) from massive unlabeled “in-the-wild” video to enable planning without action labels. Guests: No named guests; the episode is a single host-led deep dive.
16 Jan 2026 · 14m · 10 chapters
Demand estimation for differentiated products using unstructured-data proxies (images/text) in random-coefficients logit/BLP-style models; proposes a post-estimation bias correction plus proxy…
14 Jan 2026 · 14m · 6 chapters
In-context reinforcement learning with SPICE-E (SPICY) to overcome behavior-policy bias and noisy offline data using Bayesian fusion of a context evidence term with a value prior, plus posterior-UCB…
14 Jan 2026 · 12m · 7 chapters
The episode explains the “Digital Red Queen” (DRQ) algorithm, which uses an LLM (GPT-4.1 mini) to evolve Core War programs in a self-play arms race, aiming for robust generalists rather than brittle…
14 Jan 2026 · 14m · 7 chapters
The episode explains why pretrained LLMs struggle with long-context “context walls,” attributing it to rotary positional embeddings (RoPE) enabling efficient training but causing an extrapolation…
13 Jan 2026 · 12m · 8 chapters
Representation-based exploration (REPEX) for language models to enable deliberate, conceptually novel discovery instead of “sharpening” via RL; covers inference-time selection and post-training…
12 Jan 2026 · 14m · 5 chapters
NextFlow, a unified sequential (decoder-only transformer) model that combines text/image understanding and generation, aiming to match diffusion-model visual quality while retaining LLM-like…
10 Jan 2026 · 15m · 8 chapters
RelayLLM proposes “token-level collaborative decoding” so a small language model can request brief, precise help from a larger LLM during generation, avoiding all-or-nothing routing.
10 Jan 2026 · 13m · 8 chapters
Defines “hallucination” for LLMs via a unified WVP framework: hallucination is an observable mismatch between a model’s claims and an explicitly defined reference world model (W), given the model’s…
8 Jan 2026 · 12m · 4 chapters
Deep sequence models (transformers, Mamba) store knowledge in two competing ways: fast local associative memory (next-token/neighbor lookups) and slower “geometric memory” (embeddings whose dot…
8 Jan 2026 · 13m · 6 chapters
The episode argues that classical information theory (Shannon/Kolmogorov) assumes an unlimited-compute observer, so it mismeasures what AI learns.
8 Jan 2026 · 14m · 7 chapters
The paper “Diffusion Language Models Are Provably Optimal Parallel Samplers” argues diffusion language models (DLMs) can achieve provably optimal inference latency and memory for sampling, and that…
7 Jan 2026 · 12m · 9 chapters
Universal Reasoning Model (URM) for hard algorithmic reasoning benchmarks (ARC-AGI1/2, Sudoku), arguing that smaller Universal Transformer (UT) architectures beat much larger vanilla LLMs by using…
6 Jan 2026 · 14m · 10 chapters
Recursive language models (RLMs) to overcome “context rot” and extend reliable reasoning from ~200k tokens to 10M+ tokens by externalizing long prompts into a persistent workspace the model can…
6 Jan 2026 · 16m · 7 chapters
Cross-domain generalization via causal transportability for zero-shot and few-shot learning, contrasting “fast” vs “slow” adaptation.
4 Jan 2026 · 15m · 8 chapters
Argues that causal inference (interventions and counterfactuals) can be done with standard probabilistic modeling by writing a full joint probability over “observed,” “intervened,” and…
3 Jan 2026 · 12m · 4 chapters
End-to-End Test-Time Training (TTT-E2E) for long-context LLMs, aiming for full-attention-like performance up to 128K tokens while keeping constant-cost inference.
3 Jan 2026 · 14m · 7 chapters
Parallel Token Prediction (PTP) for language models to reduce slow sequential token-by-token generation latency by predicting multiple interdependent future tokens in a single forward pass.
2 Jan 2026 · 16m · 11 chapters