Awesome LLM Research Collections
  • Home
  • Papers
    • Overview
    • Attention
    • LLMs
    • Multimodal LLMs
    • Embeddings
    • SFT
    • Training
    • Reinforcement Learning
    • Agents Application
    • Vision
    • Auto-Prompt
  • Notes
  • Blogs
  • English
  • 中文

Notes

Research notes

Bilingual notes for paper readings and technical reflections around LLM research and engineering.

31Notes
8Topics
2026-09-14Latest creation
Paper Readings Technical Reflections

Browse

Paper Readings

Created 2026-09-12 · Updated 2026-09-12

The DINO Series: Architecture, Training, and Task Adaptation

A practical account of DINO, DINOv2, and DINOv3: self-distillation, local features, training steps, and losses, with a DINOv2 + SimCLR retrieval case study.

DINO Self-Supervised Learning Vision Transformer Knowledge Distillation Image Retrieval
VISION Paper Readings
Created 2026-08-17 · Updated 2026-08-22

Self-Evolving Coding Agents: Taxonomy, Feedback Loops, and Trustworthy Evolution

A structured reading of how coding agents evolve their frameworks, memory, skills and tools, models, and collaboration structures, organized by evolution object, timing, and software evidence, with an extension on validation, versioning, and rollback for trustworthy evolution.

Self-Evolving Coding Agents Coding Agents Agent Self-Evolution Software Engineering Agent Memory Agent Skills
AGENTS Paper Readings
Created 2026-08-10 · Updated 2026-08-11

SkillRL: How a Failure Becomes a Reusable Skill

A reading of SkillRL through ALFWorld's Skipping State Changes failure, tracing candidate-skill extraction, SkillBank organization, skill-aware policy training, and the boundary between paper evidence, repository behavior, and production admission controls.

SkillRL Agentic RL Reinforcement Learning Skill Learning Failure Learning Long-Horizon Agents
REINFORCEMENT-LEARNING Paper Readings
Created 2026-08-03 · Updated 2026-08-22

Kimi Linear and KDA: From Channel-Wise Forgetting to Hardware-Efficient Linear Attention

A technical reading of Kimi Linear through its recurrence, WY/UT parallel algorithm, 3:1 KDA/MLA hybrid, experiments, and public implementations.

Kimi Linear Kimi Delta Attention Linear Attention Delta Rule Long Context
ATTENTION Paper Readings
Created 2026-08-01 · Updated 2026-08-22

Kimi K3: Architecture, Training, and Million-Token Agentic RL

A plain-language explanation of Kimi K3's KDA, Block AttnRes, LatentMoE, 1M context, and nine-policy RL, with clear boundaries between paper claims and external tests.

Kimi K3 LLM Mixture of Experts Long Context Agentic RL
LLMS Paper Readings
Created 2026-07-29 · Updated 2026-08-15

Data Agents: Levels, State of the Art, and Open Problems

A lifecycle-based reading of the L0–L5 autonomy framework, tracing data agents from responsive assistants and procedural executors to supervised orchestrators while examining gaps in responsibility, governance, and evaluation.

Data Agents LLM Agents Data Management Autonomy Data Governance
AGENTS Paper Readings
Created 2026-07-28 · Updated 2026-08-22

From Qwen to Qwen3.6: The Architecture and Training Evolution of Seven Generations

A seven-generation account of how Qwen evolved across attention, MoE, data, post-training, native multimodality, and agent environments—and why Qwen3-Next is an architectural bridge rather than a separate generation.

Qwen LLM MoE Post-training Long Context
LLMS Paper Readings
Created 2026-06-30 · Updated 2026-07-01

SAPO: Replacing Hard Ratio Clipping with a Continuous Soft Gate

An analysis of SAPO through its surrogate objective, gradient weights, asymmetric temperatures, and Qwen3-VL experiments, with attention to its token-level behavior and sequence-level approximation.

SAPO Reinforcement Learning Policy Optimization GRPO GSPO Qwen3-VL
REINFORCEMENT-LEARNING Paper Readings
Created 2026-06-18 · Updated 2026-06-19

Entropy Collapse: Policy Entropy Consumption in LLM Reinforcement Learning

A note on entropy collapse in LLM reinforcement learning, covering policy entropy, the difference between SFT and RL, DAPO's Clip-Higher strategy, and covariance regularization.

Entropy Collapse Reinforcement Learning Post-training DAPO GRPO
REINFORCEMENT-LEARNING Paper Readings
Created 2026-06-15 · Updated 2026-06-15

From Qwen-VL to Qwen3-VL: Four Generations of Architecture and Training

A technical review of how four Qwen-VL generations evolved across vision-language alignment, dynamic resolution, spatiotemporal position encoding, video modeling, and deep visual fusion.

VLM Qwen Multimodal Large Language Model Position Encoding Video Understanding
MLLMS Paper Readings
Created 2026-06-15 · Updated 2026-06-15

CapRL: Stimulating Vision-Language Captioning Capabilities with Reinforcement Learning

CapRL evaluates captions through the MCQ accuracy of a vision-free LLM, turning subjective caption-quality scoring into a verifiable reward for training image-captioning models.

CapRL Reinforcement Learning RLVR Vision-Language Model Image Captioning
REINFORCEMENT-LEARNING Paper Readings

Browse

Technical Reflections

Created 2026-09-14 · Updated 2026-09-14

A Comprehensive Guide to RSI (Self-Evolution, Self-Evolving Agents)

A taxonomy of Recursive Self-Improvement (RSI) across improvement targets, evolution structures, update roles, timing, and feedback.

RSI Self-Evolving Agents Agent Harness Agent Memory Skill Evolution
AGENTS Technical Reflections
Created 2026-09-07 · Updated 2026-09-08

Query Understanding and Rewriting: Training, Retrieval, and Agent Architecture Boundaries

Intent recognition, clarification, behavior-driven rewriting, distillation, and retrieval evaluation, with evidence-dependent multi-hop training data and a discussion of Agent state management versus AIGC recaptioning.

Query Rewriting Intent Recognition Retrieval Training Data Agent Architecture
AGENTS Technical Reflections
Created 2026-09-01 · Updated 2026-09-01

From Policy Lag to Advantage Collapse: Effective Update Signals in LLM Reinforcement Learning

An analysis of ineffective LLM RL updates through rollout policies, PPO clipping, homogeneous GRPO rewards, and interventions across data, sampling, supervision, and optimization.

Reinforcement Learning PPO GRPO DAPO RLVR Advantage Estimation
REINFORCEMENT-LEARNING Technical Reflections
Created 2026-08-28 · Updated 2026-08-28

Building a Cold-Start Evaluation Dataset for AI Agents Without Production Data

A practical method for building auditable pre-launch agent eval datasets from capability maps, human-authored gold seeds, controlled mutations, state oracles, resettable environments, and release gates.

Agent Evaluation Evaluation Dataset Benchmark Design Synthetic Data Agent Safety
AGENTS Technical Reflections
Created 2026-08-27 · Updated 2026-08-27

KL Divergence: From Mathematical Derivations to LLM Reinforcement Learning

Understand the mathematical structure of KL divergence and what it controls in PPO, DPO, GRPO, and On-Policy Distillation.

KL Divergence Reinforcement Learning RLHF PPO DPO GRPO On-Policy Distillation
REINFORCEMENT-LEARNING Technical Reflections
Created 2026-08-26 · Updated 2026-08-26

Agentic RL Trajectory Data Contracts: From Harness State to Trainable Samples

A practical account of how Agentic RL preserves auditable token-level training data from model calls, context reconstruction, rollout grouping, and credit assignment.

Agentic RL Agent Harness Trajectory Data Policy Optimization Credit Assignment
REINFORCEMENT-LEARNING Technical Reflections
Created 2026-08-25 · Updated 2026-09-01

Local Credit Assignment and the Boundary of Global Correctness in OPD: Starting from a Distilled-RL Intuition

Starting from a teacher-defined dense reward intuition, this note derives OPD's reverse-KL signal, examines local agreement after a corrupted prefix, and separates the roles of KAT, BRTS, TrOPD, local support matching, and verifiers.

OPD Reinforcement Learning Distillation Credit Assignment Verifier
OPD Technical Reflections
Created 2026-08-21 · Updated 2026-08-22

Recursive Self-Improvement: From Artifacts and Harnesses to Model–Scaffold Co-evolution

A two-axis review of 2026 recursive self-improvement work, with an auditable standard for admitting self-generated updates.

recursive-self-improvement self-evolving-agents harness model-harness-coevolution
AGENTS Technical Reflections
Created 2026-08-21 · Updated 2026-08-24

What RL Actually Changes: Reasoning Patterns, Distillation, and Controllable Post-Training

A technical reflection on SFT, RLVR, distillation, and second-stage RL through the existence, selection, transfer, and control of reasoning patterns.

Reinforcement Learning SFT RLVR Distillation Reasoning
REINFORCEMENT-LEARNING Technical Reflections
Created 2026-08-19 · Updated 2026-08-28

How to Evaluate AI Agents Scientifically: Metrics, Environments, Judges, and the Regression Loop

A practical guide to AI agent evaluation covering outcome and process metrics, environment and task design, LLM-as-a-Judge, failure attribution, and dual regression.

Agent Evaluation Benchmark Design LLM-as-a-Judge Regression Testing Agent Safety
AGENTS Technical Reflections
Created 2026-08-12 · Updated 2026-08-22

From Clio to an Executable Skill: Engineering LLM-Assisted Embedding Clustering

An engineering reconstruction of Clio's semantic clustering subsystem, evaluated on 20 Newsgroups to expose the gains, costs, reproducibility limits, and missing privacy controls of LLM-assisted hierarchy building.

Clio Embedding Clustering Agent Skill Reproducibility
AGENTS Technical Reflections
Created 2026-08-10 · Updated 2026-08-22

The Benchmark Behind the Benchmark: Judge Noise and Trustworthy Agent Comparisons

An analysis of agent noise, judge variance, evidence contracts, continuous scoring, ensembles, and model comparison through a 106-task multi-judge experiment, with a clear distinction between consistency and correctness.

Agent Evaluation LLM-as-a-Judge Benchmark Statistical Reliability
AGENTS Technical Reflections
Created 2026-08-05 · Updated 2026-08-06

Treating Agents as Algorithms: Engineering Skills, CLIs, and Workflows

An engineering reflection on constraining execution variability and making multi-step agent tasks recoverable through responsibility boundaries, progressive disclosure, Gate validation, persistent state, and Workflow orchestration.

Agent Skill Design CLI Workflow Engineering
AGENTS Technical Reflections
Created 2026-07-06 · Updated 2026-07-06

Long Trajectories, Learned Values, and Adaptive Verification: The Changing Constraints of Agentic RL

A synthesis of GLM-5.2, Qwen, GenAC, OPID, and two related studies on how long-horizon agents reshape trajectory sampling, credit assignment, reward verification, and dense supervision.

Agentic RL Long-Horizon Agents Value Modeling Reward Hacking Verification
LLMS Technical Reflections
Created 2026-07-02 · Updated 2026-08-22

Keeping an SFT Gradient in DPO: From Relative Preference to a Chosen-Likelihood Anchor

A code-centered analysis of how chosen-response SFT changes DPO gradients, data and memory constraints, and the experiments needed to retain or reject the hypothesis.

DPO SFT Preference Optimization Post-training RLHF
TRAINING Technical Reflections
Created 2026-06-30 · Updated 2026-06-30

Agent Experience Learning: From Textual Reflection and Programmatic Skills to Policy Internalization

A technical synthesis of how LLM agents turn task trajectories into retrievable experience, executable skills, and parametric behavior, with an emphasis on library maintenance, evaluation, and engineering boundaries.

Agent Experience Learning Skill Learning Self-Evolving Agents
AGENTS Technical Reflections
Created 2026-06-30 · Updated 2026-07-02

Dissecting GRPO Loss in Code: Components, Negative Values, and Objective Ascent

A line-by-line analysis of importance ratios, clipping, KL penalties, and aggregation in a minimal GRPO implementation, explaining negative loss values, objective ascent, and near-zero experiments.

GRPO Reinforcement Learning Post-training Policy Optimization
LLMS Technical Reflections
Created 2026-06-29 · Updated 2026-06-29

From 1D-RoPE to Qwen's MRoPE: Frequency Allocation in Rotary Position Embeddings

Starting from a six-dimensional vector, this note derives the relative-position property of 1D-RoPE and compares blockwise MRoPE in Qwen2.5-VL with Interleaved-MRoPE in Qwen3-VL.

VLM Qwen Multimodal Large Language Model Position Encoding
MLLMS Technical Reflections
Created 2026-06-16 · Updated 2026-06-16

PPO, DPO, and GRPO: Objectives and Training Loops for LLM Alignment

A comparison of PPO, DPO, and GRPO through their objectives, advantage estimators, training loops, engineering tradeoffs, and practical boundaries.

Post-training Reinforcement Learning PPO DPO GRPO
REINFORCEMENT-LEARNING Technical Reflections
Created 2026-05-28 · Updated 2026-08-25

OPD: Capability Integration Interface in Post-training

A technical reflection on how OPD becomes a capability integration interface in post-training through Qwen3, GLM-5, MiMo-V2, and DeepSeek-V4.

OPD Reinforcement Learning Distillation
OPD Technical Reflections
  • View source
  • Report an issue