AI research01
The Missing Positional Story in LLMs
First author · NeurIPS 2026
Studying how the distance between words shapes what a language model pays attention to—and what happens when that pattern is disrupted.
Explore project Researcher & engineer / Dartmouth
Understanding intelligence.
Building things that work.
I’m Benjamin Huh. I study how AI works and build software
and devices that connect ideas to the real world.
AI researchSoftwareEmbedded systems
Selected highlights
Research, awards, and industrySoftware02
ReasonRender
Team contributor · UpScaleX Standout Team Award
Helps AI agents reuse shared task information instead of repeatedly reading the same context. Fewer tokens, less duplicated work.
Explore projectSoftware03
JWC Technologies internship
Intern · Web & embedded development
Built and launched jwctech.com during my internship and contributed to embedded firmware development.
Explore projectICML / Conference presentation
Making AI agents easier to understand.
At ICML, I presented our position paper on Research Contracts and Agent Contracts: a way to make an AI team’s goals, assumptions, and responsibilities explicit.
Position: Multi-Agent Explainability Needs Contracts before Methods
Second author · ICML 2026 Position Paper Track
01 / The project index
19 projectsIdeas, experiments, and things I’ve built.
Showing all 19 projects
AI research2026
The Missing Positional Story in LLMs
Studying how the distance between words shapes what a language model pays attention to—and what happens when that pattern is disrupted.
First author · NeurIPS 2026
Behind the project
Measured attention patterns across 11 language models and used controlled interventions to test their effect on model behavior.
Software2026
ReasonRender
Helps AI agents reuse shared task information instead of repeatedly reading the same context. Fewer tokens, less duplicated work.
Team contributor · UpScaleX Standout Team Award
Behind the project
Helped implement shared-context caching and compressed handoffs. The project received the UpScaleX Standout Team Award at the Snowflake × Beta Fund × EverMind hackathon.
Hardware2025–26
Reaction Duel
A two-player reaction game: press a button when the light turns on, then see your score travel over Bluetooth to a web leaderboard.
Firmware & full-stack development
Behind the project
Built the device firmware, Python Bluetooth bridge, and FastAPI/PostgreSQL backend. Added automated tests, deployment, monitoring, and a queue to stabilize score delivery.
AI research2026
PolicyLLM
Turns written policies into structured rules that can be checked against an AI system’s responses.
Equal contributor · ICML TAIGR workshop
Behind the project
Co-led document extraction and evaluation. My extraction module handles six document formats; the shared project connects extraction, validation, and runtime enforcement.
Hardware2024–present
Offline crowd messaging
Devices and a mobile app that help people find and message each other in large crowds, without cell service or Wi-Fi.
Collaborative project · Mobile & embedded development
Behind the project
Contributed to mobile-app design and embedded connectivity for an ongoing project.
Software2026
Fountain Pen Deal Finder
Finds fountain pens across Japanese marketplaces, compares prices, and ranks promising deals in a daily report.
Personal project
Behind the project
Built listing collection, normalization, price estimation, and deal scoring across Yahoo, Mercari, and Rakuten Rakuma.
AI research2026
Connecting sound and images
Explores how AI systems connect information across sound, images, and language.
Multimodal representation research
Behind the project
Ongoing research in multimodal representation learning.
AI research2026
Easy to Add, Hard to Erase
Tests whether changing a language model can remove stereotypes as easily as it can introduce them.
First author · StereACuLT workshop
Behind the project
Compared editing behavior across seven open-weight language models, finding that adding stereotyped responses was easier than reliably removing them.
AI research2025–present
Inside mathematical reasoning
Investigates which internal parts of a language model help it solve arithmetic and multi-step problems.
Mechanistic interpretability research
Behind the project
Ongoing research into language-model reasoning and evaluation.
AI research2026
MIRT · Choosing the right AI team
Chooses an AI model and workflow for each question while keeping cost and response time in mind.
Co-author · ICML DEMO workshop
Behind the project
Developed training and online recalibration across 33 model/workflow combinations. This continues the adaptive multi-agent routing work described in the older resume.
AI research2026
CLASP · Keeping AI within budget
Helps an existing AI router stay within cost and response-time limits as conditions change.
Co-author · ICML DEMO workshop
Behind the project
Co-developed a controller that brings routing systems back toward their operating targets without retraining the underlying router.
AI research2026
Clear contracts for AI agents
Makes AI-agent experiments easier to understand by spelling out their goals, assumptions, measurements, and responsibilities.
Second author · ICML Position Paper Track
Behind the project
Co-authored the argument for Research Contracts and Agent Contracts, with reusable templates for auditable multi-agent evaluation.
AI research2026
Translation Is Not Representation
Examines whether multilingual bias tests reflect each language, or route their judgments through English.
Co-author · StereACuLT workshop
Behind the project
A study of English-hub routing in cross-lingual bias benchmarks.
AI research2026
Lost in Translation
Studies how cultural differences can affect the way language models interpret descriptions of medical symptoms.
Co-author · StereACuLT workshop
Behind the project
Research on cross-cultural bias in LLM-assisted symptom interpretation.
AI research2026
Attention research
Studies how language models process information internally.
Attention-structure research
Behind the project
Ongoing work on understanding language-model behavior.
AI research2026
Model evaluation research
Checks whether detecting a fact inside a language model means the model actually uses that fact to answer.
Causal representation audits
Behind the project
Ongoing work on evaluating language-model representations.
SoftwareJun–Sep 2026
JWC Technologies internship
Built and launched jwctech.com during my internship and contributed to embedded firmware development.
Intern · Web & embedded development
Behind the project
Designed and developed the company’s public website with responsive pages, a contact form, and automated testing.
Hardware2024
STM32 firmware bootloader
Loads a device’s firmware and checks incoming updates before handing control to the application.
Early venture prototype
Behind the project
Built a standalone STM32L476 bootloader with separate bootloader/application memory layouts, chunked image writes, and CRC integrity checks.
AI research2026
Text and image generation
Two models built in PyTorch: one generates text, and the other learns to turn noise into handwritten-digit images.
Dartmouth coursework
Behind the project
Built reusable training, evaluation, checkpointing, and sampling pipelines for a 10.7M-parameter character transformer and an MNIST diffusion model.
No projects found.
Try a different keyword or clear your filters.
I’m a computer science master’s student at Dartmouth, working across AI research, software, and embedded systems.
In Dartmouth’s DSAIL and LISP labs, I study how language models reason, represent information, and work together. Outside the lab, I build everything from Bluetooth devices to tools for finding fountain pens.
M.S. in computer scienceDartmouth · Expected 2027
B.A. in computer scienceDartmouth · 2026