Sonde AI

AI that asks
before it guesses

We research how AI models can learn to identify gaps in their knowledge and ask the right questions — improving task completion by 69%.

Read Our Research View on GitHub

The Missing Skill in AI

AI Guesses

Current models produce plausible-sounding but wrong outputs when information is missing. They fill gaps with hallucinations instead of flagging uncertainty.

Humans Ask

Professionals know when to pause and request clarification. A doctor asks follow-up questions. A lawyer requests missing documents. This is a core skill.

We Bridge the Gap

Our methods train AI to identify knowledge gaps and ask targeted questions, bringing human-like judgment to automated systems.

Measured Impact

0.295→0.499

+69% improvement

Task completion score: without vs. with clarifying questions

30

Professional domains evaluated
law, medicine, engineering, finance...

220+

Tasks from the DOLOMITES benchmark

Evaluated on Qwen 3.5 9B using DOLOMITES benchmark with information masking

How It Works

Step 1

Mask

We systematically hide information in professional tasks, creating scenarios where AI must identify what's missing.

Step 2

Probe

Models interact with an Oracle (simulating a colleague) to ask questions and recover hidden information.

Step 3

Train

Using SFT and RL, we teach models to ask better questions — transferring question-asking skills to any model.

Open Source

gdpval-finetune

Core benchmark pipeline: masking, Oracle interaction, SFT/RL training, LLM-as-judge evaluation.

hil-bench-pipeline

SQL task evaluation: how well models identify blockers and ask for clarification.

gdpval-finetune-pipeline

Production-ready LoRA fine-tuning pipeline for GDPval tasks.

Let's Talk

Interested in our research, benchmarks, or training methods? We'd love to hear from you.