Extra Computation at Test-Time Doesn't Boost Knowledge Tasks—Yet

Test-time scaling—a technique that allocates more computing power during inference to generate longer reasoning chains—has delivered impressive gains across many AI applications. But a new study reveals a crucial blind spot: this approach doesn't reliably help when models must answer factual questions from memory alone.

Why This Matters#

Modern frontier language models tend to fabricate information (hallucinations) when asked closed-book questions where they cannot look things up. The question at the heart of this research was straightforward yet important: can we solve this problem simply by having the model think harder?

What the Research Found#

Researchers tested 14 different reasoning models using various test-time scaling strategies. The results contradicted the optimistic narrative around extended reasoning. Rather than consistently boosting accuracy, increased computation often backfired. Models that spent more compute ended up generating more hallucinations, not fewer. The team discovered that willingness to answer explained much of this dynamic—lengthier reasoning prompted models to attempt more answers, but many of those attempts were wrong.

The work also uncovered patterns resembling confirmation bias in machine learning. When given more reasoning steps, models appeared to entrench incorrect initial beliefs, elaborating on falsehoods with invented supporting details that sounded plausible.

The Information-Theoretic Angle#

The researchers offered a theoretical explanation grounded in information theory. Because test-time scaling operates as a post-processing step on a fixed model, it cannot inject new information about what the correct answer actually is. In other words, simply computing longer chains cannot overcome the fundamental limitation that the knowledge isn't reliably stored in the model's weights to begin with.

Implications#

The findings challenge a widespread assumption that raw compute at inference time can patch knowledge deficits in AI systems. For closed-book knowledge tasks, the path forward likely requires different approaches—possibly better training methods, retrieval augmentation, or architectural innovations—rather than simply letting models ruminate longer on what they already (or don't) know.