Reachability Is Not Realization: Tracing the Sources of LLM Benchmark Gains

2026-08-04

AI Summary

This study addresses the common misinterpretation that performance gains of large language models on benchmarks reflect genuine capability improvements, arguing instead that such gains often stem merely from more efficient sampling of already reachable answers rather than true problem-solving realization. The work introduces a novel distinction between reachability and realization at the individual problem level and empirically investigates it through controlled experiments that fix sampling budget, temperature, and answer format. Leveraging inference-time layer routing, answer-guided probes, MLP module ablation, and matched-budget comparisons, the authors find that random routing matches or exceeds structured search in 43 experimental settings, ablating a single MLP module recovers 68%–92% of failed cases, and training typically enhances realization without expanding—and sometimes even reducing—the upper bound of reachability, thereby demonstrating that benchmark improvements do not necessarily indicate expanded model capabilities.

Abstract

Benchmark gains are often treated as evidence of greater LLM capability. Yet the same gain can reflect different changes in model behavior. A model may reach new answers, or produce answers that were already within reach. Aggregate scores do not distinguish these changes question by question. We establish a question-level audit under fixed budgets, temperatures, and answer formats. A question is realized when the default deployment procedure produces the correct answer. A question is reachable when a specified probe finds that answer within a fixed budget. We first test whether inference-time layer routing can expand reachability. Under a matched budget, random routes match or exceed structured search in all 43 model and task settings. Answer-blind procedures retain almost none of this gain, which instead requires access to the correct answer. We then ask why reachable answers sometimes fail to appear. Across six cases spanning 0.5B to 31B, silencing one identified MLP block repairs 68 to 92 percent of a predefined failure set. We next test whether training closes the gap by expanding reachability. In five of six matched evaluations, deployed performance rises while the reachable ceiling remains flat or falls. For DAPO, the deployed score rises by 14.7 points while the reachable ceiling falls by 13.3 points. Across the settings we audit, realization and reachability therefore do not always change together. Claims of capability expansion should report both realized performance and reachability under matched evaluation conditions. Code is available at https://github.com/LiZaiyuan0619/reachability-not-realization

Problem

• reachability
• realization
• LLM benchmarking
• model capability
• evaluation gap

Innovation

• reachability
• realization
• layer routing
• MLP ablation
• benchmark evaluation