Research
The first section is papers I wrote. Below it is the reading trail that informs the work, where each line is what that paper is for here, not an abstract.
Papers
ORCID- 2026.9 · Janghoon Lee · arXiv:2609.18672
Selection Is Retrieval, Abstention Is Not: On-Device Tool Routing over 70 Korean-English Actions
Over 70 Korean and English actions on device, picking the right tool behaves like retrieval, but deciding not to call any tool does not, so the two need separate treatment.
- 2026.9 · Janghoon Lee · arXiv:2609.06129
Protocol Compression Changes Which Party Pays: Bilateral Cost in Cross-Organization LLM Agent Communication
In cross-organization LLM agent communication, compressing the protocol does not cut total cost but shifts which side pays it, measured on both parties of the exchange.
- 2026.8 · Janghoon Lee · arXiv:2608.23023
Most of the LLM Routing Gap Is Task Type: A Task-Type and Language Decomposition over a Fully Executed 14×294 Matrix, and the Reproducibility Floor It Has to Clear
A full 14-model, 294-query matrix across seven task types and three languages finds that task type explains most of the reproducible routing gap, while language recovers only a small additional share.
- 2026.8 · Janghoon Lee · arXiv:2608.13959
Repair, Not Improvement: Decomposing Constrained Decoding in Tool-Call Abstention
Separates lexical masking from termination effects in constrained tool calling and shows that constraints often repair malformed output rather than improve the model's abstention judgment.
References
- 2017 · Vaswani et al.
Attention Is All You Need
Introduces the Transformer. The attention backbone under every modern LLM and most generative stacks since.
- 2019 · Radford et al.
GPT-2
Shows unsupervised next-token training at scale yields transferable task behavior. Template for “train once, adapt with prompts.”
- 2020 · Kaplan et al.
Scaling Laws
Predictable loss vs compute, data, and parameters. The reason bigger models and careful budget allocation beat folklore.
- 2020 · Ho, Jain, Abbeel
DDPM
Denoising diffusion as a practical generative process. Foundation for image (and later video) diffusion systems.
- 2021 · Radford et al.
CLIP
Joint image-text embeddings from contrastive training. How text conditions modern image models and how we score alignment.
- 2022 · Rombach et al.
Latent Diffusion
Runs diffusion in a compressed latent space so high-res image generation is affordable. Ancestor of Stable Diffusion-class stacks we tested and rejected for natural-prompt anatomy.
- 2022 · Ho et al.
Video Diffusion
Extends diffusion to video. Establishes the denoising-over-time framing later productized at frontier scale.
- 2022 · Défossez et al.
EnCodec
Neural audio codec: speech and music as discrete tokens. Bridge from waveforms into language-model-style generation.
- 2022 · Peebles, Xie
DiT
Diffusion Transformers: replace the U-Net with a transformer over latent patches. Architecture family behind frontier image generators and ACE Step’s DiT path.
- 2023 · Lipman et al.
Flow Matching
Trains continuous normalizing flows by regressing a vector field. Cleaner, often faster alternative path to diffusion-style generation.
- 2023 · Chen, Zaharia, Zou
FrugalGPT
Prompt adaptation, approximation, and LLM cascades to cut API cost. Early framing for cheap-vs-strong routing that Redrob Eval and Console sit downstream of.
- 2024 · Ong et al.
RouteLLM
Learns routers from preference data to send easy queries to a weak model and hard ones to a strong model. Framework we tried in Eval before dropping routing at ~25% label match.
- 2024 · BerriAI
LiteLLM
Open-source multi-provider LLM gateway and routing toolkit. One of the RouteLLM-adjacent stacks exercised in the Eval routing failure.
- 2024 · Sadat et al.
APG
Adaptive Projected Guidance: fixes CFG oversaturation by projecting guidance updates. Why Tune Pro uses APG instead of plain DiT CFG on KSampler.
- 2024 · Du et al.
CosyVoice 2
Streaming, low-latency speech synthesis with strong zero-shot voice cloning. Reference point for speech models in the Redrob Speech direction.
- 2025 · Black Forest Labs et al.
FLUX.1 Kontext
Flow-matching generation and editing from Black Forest Labs. Lineage for the FLUX.2 [klein] 4B photoreal candidate we tested and rejected on typography.
- 2025 · Agrawal et al.
GEPA
Genetic-Pareto reflective prompt evolution. How Redrob Eval optimizes the cheap path before dual-eval on IN22-Gen, IndicGLUE, and GSM8K.
- 2025 · Wu et al.
Qwen-Image
Image foundation model with strong text rendering and native editing. Heavy path in Redrob Image: custom checkpoints and edit-first product work.
- 2025 · Tongyi-MAI / Z-Image Team
Z-Image
Efficient single-stream DiT image model; Z-Image-Turbo is the distilled 8-step variant. Fast path for quick and easy Redrob Image requests.
- 2026 · Sarvam AI
Sarvam
Open Indic-focused language models (30B / 105B and earlier Sarvam-1/M). Open backend family behind the routed Indic text stack in Redrob Console.
- 2026 · Gong et al.
ACE-Step 1.5
Open-source music foundation model: LM planner plus DiT, local inference, LoRA personalization. Weights and Comfy graph behind Redrob Tune Fast/Pro.