Separating training-objective effects from tool-style effects on copy-inflation

Determine how training objectives and tool-use styles independently affect copy-inflation severity and confidence-signal degradation in multi-turn search agents by evaluating controlled pairs of models or systems.

Background

The paper observes that copy-inflation severity appears to track retrieval-bag size and tool style more closely than training objective. This conclusion is based on comparisons across model families with different architectures and training procedures, but those comparisons do not isolate the relevant causal factors. Controlled pairs that hold tool style fixed while varying the training objective, or vice versa, are needed to determine the separate contributions of these factors to copying behavior and confidence miscalibration.

References

Disentangling training objective from tool style would need controlled pairs, which we leave to future work.

— Beyond Confidence: Test-Time Scaling for Multi-Turn Search Agents via Retrieval Grounding  (2608.24024 - Kook et al., 25 Aug 2026) in Appendix C, Section “Cross-dataset evidence for copy-inflation,” paragraph “Severity across model families”