Fine-tuning

HyperTool: Beyond Step-Wise Tool Calls for Tool-Augmented Agents

여러 MCP 툴 호출을 코드 블록 하나로 묶어 LLM 에이전트의 컨텍스트 낭비와 추론 단절을 동시에 해결하는 기법

ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing

도메인 파인튜닝으로 망가진 LLM 안전성을, 재학습 없이 추론 시점에 작은 안전 모델에서 빌려와 복구하는 방법.

Ultrafast machine learning on FPGAs via Kolmogorov-Arnold Networks

기존 MLP 대신 KAN 아키텍처를 FPGA에 최적화해 나노초 단위 추론과 온라인 학습을 가능하게 한 석사 논문 소개. FPGA 2026 Best Paper를 수상했고 고빈도 거래처럼 초저지연이 필수인 환경에서 주목할 만하다.

The Role of Feedback Alignment in Self-Distillation

LLM이 스스로를 가르칠 때, 피드백을 모델의 추론 흐름에 단계별로 맞추면 GRPO보다 16점 이상 수학 추론 성능이 오른다.

Can LLMs Beat Classical Hyperparameter Optimization Algorithms?

LLM 기반 하이퍼파라미터 최적화 에이전트와 CMA-ES, TPE 같은 고전 알고리즘을 직접 비교한 연구로, LLM 단독으로는 고전 방법을 이기지 못하지만 두 방법을 합친 하이브리드 'Centaur'가 최고 성능을 낸다는 결론이 나왔다.

Tiny hackable CUDA language model implementation

CUDA로 작성된 GPT(Generative Pretrained Transformer) 미니멀 구현체로, 텍스트뿐 아니라 모든 바이트 스트림을 학습할 수 있어 LLM 내부 구조를 직접 뜯어보고 싶은 개발자에게 유용하다.

My Agent Skill for Test-Driven Development

AI 에이전트가 형편없는 테스트를 작성하는 문제를 해결하기 위해, Kent Beck의 Canon TDD 원칙을 'Skill'로 만들어 에이전트에게 주입하는 방법을 공유한다. 에이전트 코딩에서 테스트 품질을 높이고 싶은 개발자에게 실용적인 접근법을 제시한다.

CS336: Language Modeling from Scratch

Stanford에서 운영하는 LLM 전 과정 구현 강의로, 토크나이저부터 데이터 수집, 트랜스포머 구현, 분산 학습, RL 기반 정렬까지 직접 코딩하며 배운다. 이론이 아닌 구현 중심이라 실제로 LLM이 어떻게 작동하는지 깊이 이해하고 싶은 개발자에게 가장 체계적인 커리큘럼 중 하나다.

1-Bit Bonsai Image 4B Image Generation for Local Devices

4B 파라미터 이미지 생성 모델의 가중치를 1비트/3값으로 극단적으로 압축해서 iPhone에서도 돌아가게 만든 모델. 7.75GB짜리 diffusion transformer를 0.93GB까지 줄였다.

LinTree: Improving LLM Reasoning with Explicitly Structured Search Histories

LLM의 추론 트레이스에 부모 포인터(parent pointer)만 추가해도 탐색 성능과 효율이 크게 올라간다.

Token-Level Generalization in LoRA Adapter Backdoors: Attack Characterization and Behavioral Detection

HuggingFace에서 다운받는 LoRA 어댑터에 백도어를 숨길 수 있고, 이를 탐지하는 방법도 있다.

Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases

LLM이 자기 자신의 RLHF 학습 과정을 조작해 편향을 증폭시키는 구조적 취약점을 발견했다.

A sleep-like consolidation mechanism for LLMs

LLM이 긴 컨텍스트를 처리할 때 발생하는 Attention 비용 문제를 해결하기 위해, 사람의 수면처럼 주기적으로 컨텍스트를 fast weight에 압축·저장하는 새로운 메커니즘을 제안한 논문이다.

PopuLoRA: Co-Evolving LLM Populations for Reasoning Self- Play

단일 모델 self-play의 고질적 문제인 '난이도 붕괴'를 교사-학생 LoRA 집단의 공진화(co-evolution)로 해결한 연구로, 수학·코드 벤치마크 다수에서 baseline을 뛰어넘었다.

Formal Verification Gates for AI Coding Loops

AI가 생성한 코드에서 보안 불변식(invariant)을 지키게 하려면 프롬프트 지시보다 타입 시스템 같은 구조적 제약이 훨씬 효과적이라는 주장과 구현 방법을 소개한다.

FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast

파인튜닝 없이 여러 AI 에이전트가 실패 경험을 공유하며 집단지성으로 메모리를 진화시키는 프레임워크

MetaBackdoor: Exploiting Positional Encoding as a Backdoor Attack Surface in LLMs

입력 텍스트는 멀쩡한데 입력 길이만으로 LLM 백도어가 발동되는 새로운 공격 기법 발견.

Negation Neglect: When models fail to learn negations in training

"이건 가짜입니다"라고 수천 번 경고해도, 그 문서로 파인튜닝하면 모델은 내용을 사실로 믿어버린다.

Training an LLM in Swift, Part 1: Taking matrix mult from Gflop/s to Tflop/s

Apple Silicon에서 Swift로 직접 행렬 곱셈 커널을 구현하며 CPU, SIMD, AMX, GPU(Metal)를 단계별로 최적화해 Gflop/s에서 Tflop/s 수준까지 성능을 높이는 과정을 상세히 설명한 글이다. 프레임워크 없이 LLM 학습의 핵심 연산을 밑바닥부터 구현하고 싶은 개발자에게 Apple Silicon의 성능 한계를 체감할 수 있는 드문 자료다.

Conceptors for Semantic Steering

LLM의 hidden state에 행렬 기반 'conceptor'를 끼워서 감정·정치성향·우울 같은 개념을 재학습 없이 정밀하게 조종하는 방법

Refusal in Language Models Is Mediated by a Single Direction

Open-source chat models encode safety as a single vector direction, and removing it disables safety fine-tuning.

Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library

PyTorch Lightning packages 2.6.2 and 2.6.3 delivered credential-stealing malware via a supply chain attack.

Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

Fine-tuning even safety-aligned LLMs can bypass safeguards and reproduce copyrighted text verbatim, revealing prompt filtering alone isn't enough to prevent copyright infringement.

Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers

Even safety-evaluated LLMs exhibit hazardous behavior when triggered by specific contextual cues.

EvanFlow – A TDD driven feedback loop for Claude Code

EvanFlow automates code brainstorming, TDD, and validation in Claude Code with 16 skills triggered by a single prompt.

HarDBench: A Benchmark for Draft-Based Co-Authoring Jailbreak Attacks for Safe Human-LLM Collaborative Writing

LLM-Refine benchmark reveals large language models readily complete instructions for building explosives.

FUSE: Ensembling Verifiers with Zero Labeled Data

FUSE automatically ensembles multiple LLM verification models without ground truth labels, achieving Best-of-N performance comparable to semi-supervised learning.

Show HN: MacMind – A transformer neural network in HyperCard on a 1989 Macintosh

This is an educational project implementing a single-layer Transformer with 1,216 parameters in the scripting language HyperTalk (1987) and training it on a real Macintosh SE/30. It demonstrates that the core mathematics of modern LLMs works the same on hardware from 30 years ago.

One Token Away from Collapse: The Fragility of Instruction-Tuned Helpfulness

We discovered that LLM responses can shrink by up to 48% with a single instruction: "Don't use commas".

HiL-Bench (Human-in-Loop Benchmark): Do Agents Know When to Ask for Help?

A benchmark for measuring an AI coding agent's ability to determine when to ask humans when given incomplete specifications.

Dynamic Context Evolution for Scalable Synthetic Data Generation

A framework that completely eliminates duplication and repetition in large-scale synthetic data generation with LLMs using three mechanisms (VTS + Semantic Memory + Adaptive Prompt).

MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU

Introducing MegaTrain, a system that leverages CPU memory as the primary storage and utilizes the GPU solely as a compute engine, enabling full-precision training of 120B parameter models with just a single H200 GPU.

Show HN: I built a tiny LLM to demystify how language models work

This educational project allows you to build a mini LLM with 8.7 million parameters, trained on a Guppy fish character, from scratch in just 5 minutes using a single Colab notebook, focusing on demystifying the black box nature of LLMs.

Show HN: Real-time AI (audio/video in, voice out) on an M3 Pro with Gemma E2B

We open-sourced a real-time multimodal AI speech and video conversation system that runs completely locally on Apple Silicon M3 Pro without the internet. It is attracting attention for its ability to handle speech recognition, video understanding, and TTS simultaneously without cloud costs.

Nanocode: The best Claude Code that $200 can buy in pure JAX on TPUs

An open-source library that allows you to train a 1.3B parameter coding agent model from scratch on a $200 (approximately 270,000 KRW) TPU, following Anthropic's Constitutional AI approach. It can serve as a hands-on reference for developers who want to directly understand the entire AI training pipeline.

Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs

PrismML has released the Bonsai LLM series (8B/4B/1.7B) based on 1-bit weights, claiming 14x memory reduction, 8x speed improvement, and 5x energy savings compared to conventional 16-bit models, while achieving comparable benchmark performance.

Ollama is now powered by MLX on Apple Silicon in preview

Ollama has switched its inference backend on Apple Silicon from llama.cpp to Apple's MLX framework, delivering up to nearly 2x faster inference speeds. On M5 chips, it also leverages the GPU Neural Accelerator, bringing meaningful performance gains to coding agent workflows.

Hamilton-Jacobi-Bellman Equation: Reinforcement Learning and Diffusion Models

A math blog post showing how 1840s physics equations connect modern RL and Diffusion Models, explaining that continuous-time RL and generative model training are two faces of the same optimal control problem.

From 300KB to 69KB per Token: How LLM Architectures Solve the KV Cache Problem

A breakdown of how LLM KV Cache architecture has evolved from GPT-2 to DeepSeek V3, comparing per-token memory costs across architectures as they dropped from 300KB to 69KB.

CERN uses ultra-compact AI models on FPGAs for real-time LHC data filtering

CERN uses a 'hardware-first' inference approach at the LHC by burning PyTorch/TensorFlow models directly into FPGAs to filter hundreds of terabytes of collision data per second at nanosecond latency — a radical departure from conventional GPU/TPU-based AI.

If you don't opt out by Apr 24 GitHub will train on your private repos

Starting April 24, GitHub changed its policy to use Copilot users' private repo interaction data for AI training by default. You need to know exactly where the opt-out link is and what data is actually in scope.

Running Claude Code fully offline on a MacBook — no API key, no cloud, 17s per task

A post sharing how to run Claude Code fully offline on a MacBook by connecting it to a local LLM without an API key or cloud, useful for developers who want to use an AI coding assistant at no cost.

TurboQuant: Redefining AI efficiency with extreme compression

Google Research 2-stage vector compression — PolarQuant + QJL achieves 6x KV cache reduction with zero accuracy loss and 8x attention speedup on H100 GPUs

Hypura – A storage-tier-aware LLM inference scheduler for Apple Silicon

A Rust-based open-source project that intelligently distributes LLM models across GPU, RAM, and NVMe when they exceed your Mac's physical memory, enabling models that crash llama.cpp with OOM errors to actually run.

LLM Neuroanatomy II: Modern LLM Hacking and Hints of a Universal Language?

A training-free technique (RYS) that duplicates Transformer layers works across all modern LLMs — and reveals that internal representations converge toward a "universal language" independent of human language.

SPA: A Simple but Tough-to-Beat Baseline for Knowledge Injection

7 cognitively-grounded prompt templates turn a small domain corpus into massive synthetic training data — and outperforms complex RL/multi-stage approaches at knowledge injection.

[R] Doc-to-LoRA: Learning to Instantly Internalize Contexts from Sakana AI

Sakana AI D2L — hypernetwork generates LoRA adapter from a document in a single forward pass, sub-second latency, extends context window 5x beyond base model capacity

NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute

Achieved 10x data efficiency in a few weeks — training a 1.8B parameter model ensemble on only 100M tokens to match the performance of 1B token training. An approach for preparing for a future where compute is abundant but data is the bottleneck.

Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding

Extracting the implicit 3D spatial knowledge learned by video generation models (Wan2.1) to boost MLLM spatial reasoning ability.

F2LLM-v2: Inclusive, Performant, and Efficient Embeddings for a Multilingual World

Multilingual embeddings supporting 200 languages without English bias that outperform Qwen3-Embedding at smaller sizes.

Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation

NVIDIA's recipe for training a 30B MoE open-source reasoning model that won gold medals at IMO and IOI 2025.

Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster

An experiment report: give Claude Code 16 GPUs and it runs 910 experiments in 8 hours, achieves a 2.87% improvement in validation loss, and develops its own strategy for leveraging a mixed H100/H200 hardware pool.

Context Bootstrapped Reinforcement Learning

Gradually injecting few-shot examples early in RL training then slowly removing them lets the model internalize reasoning patterns on its own.

Memento-Skills: Let Agents Design Agents

A system where agents self-evolve by accumulating executable 'Skill' files as external memory, without touching LLM parameters

Unified Spatio-Temporal Token Scoring for Efficient Video VLMs

A lightweight token pruning module that cuts 50% of visual tokens in video AI models with only 0.7% performance loss

Loc3R-VLM: Language-based Localization and 3D Reasoning with Vision-Language Models

A framework that gives VLMs 3D spatial understanding and self-localization using only regular monocular video

EchoGen: Cycle-Consistent Learning for Unified Layout-Image Generation and Understanding

A study where image layout generation and image understanding (grounding) help each other within a single model, improving both tasks

Online Experiential Learning for Language Models

An LLM framework that keeps learning from real-world usage after deployment — no reward functions, no human labeling needed.

LLM Architecture Gallery

Dr. Sebastian Raschka put together a one-page gallery with architecture diagrams and key specs for dozens of major LLMs — Llama, DeepSeek, Qwen, Gemma and more — so you can compare design decisions at a glance.

Tree Search Distillation for Language Models Using PPO

Like AlphaZero, this trains LLMs by using MCTS to find stronger reasoning paths and then uses PPO to distill those paths back into the model.