Core
Model reliability and confidence estimation
Uncertainty quantification for language models: estimating, answer by answer, how likely an output is to be correct. This is the center of our work.
Research
Research first, honestly reported.
Kanseko began in model distillation research. That work led to the question at the center of what we do now: can a language model's reliability be measured one answer at a time?
Research areas
What we work on
Active
Inference-time methods
Ways to improve model accuracy without retraining.
Origin
Knowledge distillation
Transferring capabilities from large models to small ones. This is where Kanseko started, and it remains part of our ongoing research.
Discipline
Evaluation methodology
Pre-registered experiments, controlled baselines, and scoring only on answers the system never saw during fitting.
Two questions
Is this answer wrong? Is this answer right?
Error detection, also called failure prediction or hallucination detection, asks which answers are most likely to be wrong, so that checking effort goes where the errors actually are.
Correctness certification, also called selective prediction, asks the opposite question: which answers can be trusted confidently enough to skip review entirely.
The two are complementary, and we study both. Our research also covers generation failures: answers that fail to produce a usable result.
Methodology
We try hard not to fool ourselves.
How reliable a model is depends on the model and the task, so a single headline number rarely tells the real story. We look at how accuracy changes as more answers are accepted without review, and report the whole curve.
- Pre-registered experiments: we decide what we're testing before we run it.
- Controlled baselines, so improvements are measured against something real.
- Scoring only on answers the system never saw during fitting.
Early results
What we're seeing so far
On its own, the model we tested answers 65.4% of questions correctly. Every result below is scored on answers the system never saw during fitting.
of the model's most confident 20% of answers were correct, using the Standard check.
precision when auto-approving half of all answers with the Enhanced check, against the model's 65.4% baseline.
of errors caught by flagging 35% of answers with the Enhanced check, raising accuracy on the rest to 91.1%.
Correctness
Auto-approving right answers
The system accepts its most confident answers and sends the rest to review. Precision is the share of accepted answers that were actually right, and "review wrong" is the share of the review pile that was actually wrong. Each row is one operating point, so the two sides add up to every answer.
Standard uses just one answer per question. Enhanced uses up to three.
| Accepted | Precision | Sent to review | Review wrong |
|---|---|---|---|
| 11% | 100% | 89% | 39% |
| 20% | 100% | 80% | 43% |
| 31% | 93% | 69% | 47% |
| 41% | 90% | 59% | 52% |
| 50% | 86% | 50% | 55% |
| 62% | 83% | 38% | 62% |
| 72% | 80% | 28% | 73% |
| 80% | 79% | 20% | 88% |
| 89% | 73% | 11% | 96% |
| Accepted | Precision | Sent to review | Review wrong |
|---|---|---|---|
| 12% | 100% | 88% | 39% |
| 21% | 100% | 79% | 44% |
| 32% | 99% | 68% | 50% |
| 38% | 99% | 62% | 55% |
| 50% | 97% | 50% | 66% |
| 61% | 92% | 39% | 76% |
| 72% | 85% | 28% | 84% |
| 80% | 80% | 20% | 93% |
| 89% | 73% | 11% | 100% |
Error detection
Catching wrong answers
Flag the answers most likely to be wrong. Precision is the share of flagged answers that really were wrong.
| Share flagged | Check | Precision | Errors caught | Accuracy of the rest |
|---|---|---|---|---|
| 5.6% | Standard | 100% | 16% | 69.3% |
| 9.2% | Enhanced | 100% | 27% | 72.0% |
| 13% | Enhanced | 88% | 33% | 73.3% |
| 25% | Enhanced | 86% | 63% | 82.8% |
| 35% | Enhanced | 82% | 83% | 91.1% |
| 53% | Enhanced | 63% | 97% | 97.5% |
Three-way triage
Accept, review or reject
With decision lines set in advance and then applied, each answer is sorted into one of three outcomes. With the Enhanced check:
Accept
60%
of answers accepted, and 92% of those were right.
Reject
13%
of answers rejected, and 92% of those were wrong.
Review
27%
sent to review, where two-thirds were wrong, so review effort lands where errors are.
With more cautious settings, the system accepts 37% of answers at 96% correct with Enhanced, or 21% at 95% correct with Standard.
Benchmarks
Model Benchmarks
Coming soon.
Publications
Papers and write-ups
Publications coming soon.
Contact
Interested in our work?
For more information about Kanseko and our research, email us.
contact@kanseko.com