Topic
Qwen3
16 episodes
-
Overview: Active vs Total Parameters
We finally slow down on active vs total parameters, because we keep throwing the phrase around like it explains itself. This is us making the difference click: what a model stores versus what it actually uses when it answers.
-
Agentic Rl
Cameron Wolfe's 'Agentic RL' argues that training LLMs for agentic work requires shifting from single-turn reasoning to multi-turn trajectory optimization, where the harness (tools, environment, memory) becomes part of the RL loop itself. The central claim is that standard post-training methods fail on long-horizon tasks because they don't account for environment state changes across steps, necessitating new rollout infrastructures and stability techniques like PPO over GRPO for variable-length traces.
-
When is Your LLM Steerable?
A new paper investigates whether you can predict if a steering attempt will succeed by looking at the model's hidden states from just the first few tokens—before decoding the full response. The authors build ASTEER, a dataset of 1.4M steered generations labeled for success/failure across 150 concepts, then train a GBDT classifier on early hidden-state features to predict three outcomes: under-steer, success, or over-steer. The predictor hits ~0.7 macro-F1 on unseen concepts and cuts the cost of finding good steering strengths without expensive full rollouts.
-
Smaller Models are Natural Explorers for Policy Level Diversity in GRPO
Justy and Cody discuss a paper proposing S two L P O, a small-to-large rollout strategy for G R P O that uses smaller same-family models as structured explorers for reasoning model training.
-
End to End Context Compression at Scale
Justy and Cody dig into Latent Context Language Models (LCLMs) — encoder-decoder compressors that shrink long prompts into short latent sequences, cutting memory and latency at ratios up to 1:16 while staying competitive on accuracy. They cover the architecture search, the training recipe, the agent use-case, and what production deployment actually looks like.
-
Rethinking Continual Experience Internalization for Self Evolving LLM Agents
Jingwen Chen et al. diagnose why iterative experience internalization fails in LLMs and prescribe a three-part fix—principle-level granularity, step-wise injection, off-policy context-distillation—that turns capability collapse into compounding improvement.
-
MemTrain: Self Supervised Context Memory Training
Self-supervised framework MemTrain improves LLM context memory by training on unlabeled Wikipedia with coupled proxy tasks—masked reconstruction and memory recall—using GRPO. Achieves up to 17.67-point gains on long-horizon reasoning without task-specific labels.
-
LongTraceRL: Learning Long Context Reasoning from Search Agent Trajectories with Rubric Rewards
Justy and Cody unpack LongTraceRL, a paper that trains long-context reasoning models using realistic search-agent distractors and entity-level rubric rewards, with a short look at what would make it shippable.
-
5 Small Language Models for Agentic Tool Calling KDnuggets
Small language models are gaining ground on a critical frontier benchmark: tool calling. This episode looks at five compact, open-weight models that can route to APIs, format JSON arguments, and run multi-step agentic workflows without requiring a data center. Cody and Justy debate whether the gap between small and frontier models is closing fast enough to matter for real shipping teams.
-
Agent Evals
Justy and Cody dig into Cameron Wolfe’s argument that agent evals need to move from static benchmark thinking to realistic harnesses that test autonomy, tool use, recovery, and long-horizon behavior. They get specific about the agentic loop, why tool-call correctness is only part of the story, and where outcome-based evals can hide ugly behavior. Cody mostly buys the technical framing, with caveats about overfitting to harnesses and the difficulty of defining ground truth trajectories. Justy keeps pulling it back to who actually needs this now: teams shipping coding, workflow, or other higher-stakes agents where a demo is not the same as reliability.
-
DV World: Benchmarking Data Visualization Agents in Real World Scenarios
Justy and Cody dig into DV-World, a new benchmark from a multi-institution research team that stress-tests AI data visualization agents on real-world tasks — spreadsheet manipulation, cross-framework chart evolution, and handling ambiguous user intent. Even the best models top out around 50%, which tells you a lot about where the gap actually is.
-
LongAct: Harnessing Intrinsic Activation Patterns for Long Context Reinforcement Learning
Justy and Cody dig into LongAct, a paper about making long-context RL work better by updating only the attention weights tied to unusually large query and key activations. They unpack why that matters for long docs, agents, and multi-step reasoning, how the saliency-guided sparse updates map activation outliers back to specific weight rows, and why the reported gains across LongBench v2, RULER, and multiple RL algorithms suggest this could be more than a lab curiosity.
-
Embarrassingly Simple Self Distillation Improves Code Generation
Apple researchers developed Simple Self-Distillation (SSD), a technique that improves code generation models by fine-tuning them on their own raw outputs—no verification needed. The method improved Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench by reshaping token distributions to balance precision and exploration in code generation.
-
DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning
DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning Yicheng Chen 1,2 , Zerun Ma 2 , Xinchen Xie 2 , Yining Li 2† , Kai Chen 2† 1 Fudan University 2 Shanghai AI Laboratory Github : https://github.com/yichengchen24/DataChef Abstract In the current landscape of Large Language Models (LLMs), the curation of large-scale, high-quality training data is a primary driver of model performance. A key lever is the data recipe , which comprises a data processing pipeline to transform raw sources into training corpora.
-
We Got Claude to Fine Tune an Open Source LLM
The recent development allowing Claude to fine-tune open-source language models marks a significant step in democratizing AI training. It simplifies the complex process of model training, making it accessible to more users and applications, ultimately driving innovation in various sectors.
-
LLM Evaluation 4 Approaches
Understanding the 4 Main Approaches to LLM Evaluation (From Scratch) Multiple-Choice Benchmarks, Verifiers, Leaderboards, and LLM Judges with Code Examples Sebastian Raschka, PhD Oct 05, 2025 319 25 30 Share How do we actually evaluate LLMs? It’s a simple question, but one that tends to open up a much bigger discussion.