Join us

Lab notes

Subscribe

2026

A ScrambleToolBench episode: an agent probes scrambled commands, follows responses to a passcode, submits it and reuses tool knowledge in the next task
September 2026 · Vernon Toh

ScrambleToolBench: the next step is already in the map

An agent discovers what its tools do. Their names change, and it starts searching again—even when its earlier observations point to the next call.

Read note
A successful MNIST-PRO episode: four glimpses along a handwritten zero, ending with the correct answer at step 9
September 2026 · Vernon Toh

MNIST-PRO: seeing the strokes is not enough

When an agent gets a digit wrong, did it miss the evidence or struggle to put the glimpses together? We use MNIST-PRO to tell these failures apart.

Read note
IDEAgent quality-diversity search framework
August 2026 · Varun Gumma

IDEAgent: Agentic Quality-Diversity Search

IDEAgent searches for a diverse set of candidate research ideas and records how each candidate changes during refinement.

Read note
PODS oscillatory data-volume scheduling overview
May 2026 · Suorong Yang

Toward Efficient Data-Centric Training, Part II

PODS changes how much data is used at each stage of training instead of fixing one selection ratio.

Read note
Data Agent training-loop overview
May 2026 · Suorong Yang

Toward Efficient Data-Centric Training, Part I

Data Agent updates its data-selection policy as the target model trains.

Read note
delta-mem architecture overview
May 2026 · Jingdi Lei

δ-mem: Efficient Online Memory for LLMs

δ-mem keeps a compact state during inference and uses it to modify attention in a frozen language model.

Read note