A Verified Guess Is Not a Reversed Fact: What Test-Time Thinking Repairs in the Latent Generalization of Fine-Tuned Knowledge, Why Its Gains Follow the Forward Paths a Query Leaves Open, and the Candidate-Pool Controls That Would Separate Inversion F
(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research A language model fine-tuned on new facts often cannot use them the way it uses the same facts placed in its context: it fails to answer reversed questions and to chain two trained facts into a conclusion. A 2026 study trained a reasoning model with reinforcement learning to think over its own fine-tuned knowledge, reported that thinking repairs many of these failures and transfers to knowledge on which no reinforcement learning was run, and found one exception: pure reversals, where thinking does not invert a stored fact but reaches above-chance accuracy by proposing candidates and checking each in the trained direction. In this analysis the author reviews and synthesises published results to show that the exception is the rule seen from its other side. Test-time thinking can traverse an association only in the direction in which it was stored, so its gains track whether the query leaves a forward path to the answer. On the author's arithmetic from the study's own figure, thinking closes 56 to 88 percent of the gap between fine-tuning and in-context learning on the three holdouts where such a path exists, and 10 to 31 percent on the two where the answer can be reached only through a reverse key. Where no path exists, a reasoning model can still turn inversion into search, but search succeeds only as far as three things allow: a candidate pool small and salient enough to enumerate, coverage of that pool within the trace, and a verifier that does not accept wrong candidates. The published record bears on each: verification of facts is learned before generation, yet it can affirm both an old and a new answer, and longer reasoning raises hallucination when it raises the attempt rate. The paper consolidates the measured values in Table 1 and Figure 1, states the decision structure as Algorithm 1, and specifies the pool-size, direction-swap and oracle-verifier controls that would separate inversion from search.
Authors
- Pranay M. Mahendrakar (ORCID: https://orcid.org/0009-0003-7224-029X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24247220
- Primary Topic
- Topic Modeling
- Type
- preprint