Duration reasoning: from comparison to numerical answers | Mindtro

Adding a task that teaches duration ratios improved how the model compares durations. That improvement was not sufficient to produce accurate numerical answers for the user.

Research question

Does also teaching the model ratios between durations help it identify which duration is longer and estimate its value?

Method

Standard duration training was compared with training that gave a small additional weight to learning duration ratios. Each approach used three random initial settings. Tests on intervals outside the training conditions measured correct ordering and whether numerical answers fell within the allowed error margin.

Results

Correct duration ordering increased from 74.49% to 76.82%, a gain of 2.33 percentage points. Errors in duration ratios and generated duration values also decreased. However, the share of numerical answers within the accepted error margin remained 0% on this test. The learned duration relationships improved, but the reliable-answer objective was not met.

Limitations

This is not a completed system for answering duration questions. Further work must align predictions with real durations and limit answers when the model is uncertain.

About this report

These are technical research reports, not peer-reviewed publications. Each report identifies its source, measurement conditions and limitations. The evidence section explains what can be downloaded or reproduced.

Technical notes and source records

Experiment: EXP008d; auxiliary ratio-task weight: 0.02. Ratio error (log10-MAE) fell from 0.2254 to 0.1565; generated-duration error fell from 0.6379 to 0.4934. This is mean absolute error on a base-10 logarithmic scale; lower is better. The source metric “renderer tolerant accuracy” measures answers within the specified error margin.

The file names and SHA-256 fingerprints identify the exact source versions used to prepare this report.

EXP008d-3seed.mdSHA-256: 0b21f6de00d4fe1ceb90c4841af86e1a1ac440ae6b9230029307d0a39f7f8bad
EXP008d-renderer-3seed.mdSHA-256: 85415d6a0b0d6df104964ec591bc28dfb9fbea17b1eb7a723a49363e2627380f
EXP008d-chronoscale-3seed.mdSHA-256: 75fdf6dc0471871082efb7142b600d539d423a83a4d7bcfa5e42498c1d2a4c00

Request the research materials →