BITLABS東京 AI研究開発

Research

Research that makes AI systems more usable in production.

At BitLabs, we actively research AI and build PoCs to test model behavior, inference design, and reliability before production decisions are made.

Public Lab Notes

Research and PoC work help us separate promising ideas from deployable systems.

Some research and PoC work stays private because of confidentiality constraints.

Pre-training

LLM and SLM training programs

Focus
Data quality, objective design, and training efficiency.
Method
Review datasets, block weak runs early, and align experiments to target use cases.
Signal
Training work that stays connected to product value.

Fine-tuning

Adaptation with evaluation attached

Focus
Domain behavior, tool use, and policy alignment.
Method
Run fine-tuning beside regression checks and task scorecards.
Signal
Model changes can be measured before release.

Inference

Serving and inference stack design

Focus
Latency, GPU efficiency, and deployment limits.
Method
Study batching, KV cache, memory pressure, and traffic patterns together.
Signal
More predictable performance in production.

Agents

Agent reliability and control

Focus
Tool boundaries, review paths, and failure handling.
Method
Replay traces, test permissions, and measure recovery behavior.
Signal
Agents stay useful without losing operational control.

Model and Inference Research

We study both the model program and the serving layer.

Check 01

Which data quality signals should stop a training run early.

Check 02

Which evaluation results actually predict usefulness in the target workflow.

Check 03

How serving limits should shape model choice before product work starts.

Reliability

Research matters because AI systems need evidence before they earn production trust.

Check 01

Evaluation across quality, latency, safety, and controllability.

Check 02

Reliability testing for tool use, planning behavior, and escalation paths.

Check 03

Release criteria aligned to governance and deployment requirements.