Open Models Almanac
Methodsdata curationsynthetic data

single source · data curation

Backfilling missing reasoning between tool calls

Backfilling adds reasoning to tool calls in trajectories where it is missing.1 A teacher model writes reasoning that leads to the already-recorded tool call, kept only if a second LLM call derives that same call from the reasoning alone.2

Trade-off It prevents tool-calling models from repeating failed calls instead of recovering, and lets the agent track its plan and handle exceptions.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. we add reasoning to tool calls where it is missing, in a step we call backfilling — tech report
  2. A teacher model writes reasoning that leads to the recorded tool call, which stays unchanged. The reasoning is kept only if a second LLM call derives the same tool call from it alone — tech report
  3. T ool-calling models tend to repeat a failed call instead of recovering from it (Junhao Su et al. 2026), and reasoning between actions lets an agent track its plan and handle exceptions — tech report

Also called backfilling, Tool-call reasoning backfilling, reasoning backfilling.

Filed under data curation › synthetic data