Search papers, labs, and topics across Lattice.
This study investigates the efficacy of candidate conditioning in test-time aggregation for AI models, particularly when all candidates are incorrect. By introducing a candidate-free control, the authors demonstrate that while conditioning on multiple correct candidates significantly improves accuracy, it conversely decreases accuracy when all candidates are wrong. The findings reveal a nuanced relationship between candidate context and performance, suggesting that fresh solving can outperform conditioned responses in certain scenarios.
Candidate conditioning can boost accuracy by nearly 30% when multiple correct answers are present, but surprisingly, it can hinder performance when all candidates are incorrect. WHY_IT MATTERS: This insight challenges existing assumptions about candidate conditioning, potentially reshaping strategies for efficient test-time reasoning in AI systems.
When every candidate is wrong, correct-candidate selection is unavailable, yet the aggregation call can still solve the problem afresh. A correct aggregate answer may therefore reflect recombination, fresh solving, or both. For efficient test-time reasoning, the relevant question is whether candidate context adds value beyond the additional generation pass. We introduce the missing candidate-free control under the same maximum output-token allowance and stratify by the number of correct candidates. Across AIME-2025 and HMMT-2025 with Qwen3-4B, candidate conditioning improves accuracy when multiple candidates are correct ($\Delta_{\mathrm{cand}}$(c2+) = +0.290), lowers accuracy when every candidate is wrong ($\Delta_{\mathrm{cand}}$(c0) = -0.123), and remains unresolved in the one-correct regime. The c2+ and c0 conclusions survive a conservative correction for the adaptive two-benchmark procedure. Under this counterfactual, the interpretation of all-wrong recovery reverses at this scale: conditioning on an all-wrong candidate pool lowers accuracy relative to a fresh solve. Original-format matching and placebo results characterize the failures descriptively but leave their mechanism unresolved. Within a separate structured intervention, explicit answer fields causally steer outputs toward their values; masking yields no measurable accuracy improvement, and equivalence with the original format was not established. The evidence is limited to one Qwen3-4B family, two mathematics benchmarks, first-answer-truncated candidate fragments, and single-pass prompted aggregation.
