In her breakdown of Gekhman et al.'s "Thinking to Recall", Vivedha Elango adds her own experiment on top of the summary — and that experiment is what makes the piece worth reading. The paper's central claim: reasoning does more than handle "hard problems" — it widens the boundary of what a model can recall from its own weights. 90% of SimpleQA-Verified questions are single-hop factual lookups, and thinking still helps. There are two mechanisms, and they're separable experimentally: a computational buffer (the filler "Let me think" ×200 raises accuracy — extra forward passes let the hidden state refine itself) and factual priming (the model generates adjacent facts and walks itself to the target — nine kings of Nepal lead to the tenth).
The advice that didn't survive its test
The most valuable thing in the article isn't the result, it's the discipline. The author extrapolated the finding to RAG: a facts-first pipeline — retrieve, compress into a fact list, answer from the list. She tested it before publishing — and the extrapolation didn't survive: identical accuracy (82.4% vs 82.4%) at ~6.5× the latency. The explanation is honest: priming unlocks knowledge locked in the weights; in RAG the answer already sits in the retrieved context, there's nothing to unlock, and what remains is an extra lossy pass. On Medium, where advice ships without a run, the negative result is a genre in desperately short supply. Mechanisms don't transfer between regimes for free — that meta-lesson is worth more than any percentage gain.
What I'm taking for local models
- Weaker models gain more. The reasoning gain is ordered inversely to model capability: Qwen3-32B gains more than Gemini 2.5 Flash, which gains more than Pro. A small model has more "locked" knowledge. For a local deployment on your own hardware this changes the math: a thinking budget on a 30B model pays off where a frontier model burns it for nothing.
- Decide by metrics, not by taste. Low top-1 with high pass@k — the model "knows but can't reach", thinking will help. Both low — the knowledge isn't in the weights, you need retrieval, and thinking tokens burn budget for nothing.
- Facts-first prompting for closed-book: the first call lists the relevant facts, the second answers from the list with reasoning OFF. Close to full thinking at a fraction of the compute.
The dark side matters more than the benchmarks
A hallucination in an intermediate fact causally drags the final answer down: regression slope 0.84–0.86; clean traces score 71.1% against 32.2% for contaminated ones. A reasoning model wraps a wrong answer in a convincing chain of intermediate "facts" — false confidence instead of honest ignorance. In a multi-step agent this compounds: step 2's hallucination becomes step 3's priming context. The cure without retraining is trace selection ("Only Facts" +8.2%, "Only Correct Facts" +12.2% relative), with one hard requirement: the verifier must be able to abstain — a hallucinating verifier is worse than none.
Bottom line
Thinking is a tool with a measurable zone of applicability, not a "make it smarter" toggle. The boundary runs along pass@k, the price along trace cleanliness, and advice — your own included — lives until the first run.