Two new papers on getting a model to say how sure it is
SymboUQ proposes symbolic uncertainty quantification for spatial reasoning; TrAC uses trace-conditioned answer consistency for efficient uncertainty estimates. Both attack the gap between an answer and a calibrated confidence in it.
Two papers from the same group landed in the August listings: SymboUQ, on symbolic uncertainty quantification for spatial reasoning, and TrAC, on trace-conditioned answer consistency as an efficient route to uncertainty estimates.
Uncertainty quantification is the unglamorous problem that determines whether any of this is usable in a serious setting. A model that is right 90% of the time and cannot tell you which 90% is difficult to build a process around. One that is right 80% of the time and reliably flags its own doubt is straightforwardly useful — you route the flagged cases to a person.
The efficiency framing in TrAC is the practical part. The established way to estimate confidence is to sample the same question many times and measure agreement, which multiplies cost by the sample count. Anything that gets a calibrated number from fewer passes changes what is affordable in production, not just what is possible in a paper.
SymboUQ's narrower domain is a feature rather than a limitation. Spatial reasoning has structure a symbolic method can exploit — geometry constrains what answers are even coherent. General-purpose uncertainty over free text has no such scaffolding, which is why it remains hard.
Both belong to the same turn in the literature as the work on latent reasoning: less interest in making models produce more text, more interest in knowing what the text is worth.
arXiv — Artificial Intelligence, August 2026 listing → · arXiv — Machine Learning, August 2026 listing →