Think harder, same answer: reasoning buys confidence, not correctness
In a preprint, shopping agents that reasoned longer grew more confident without getting more accurate. Two more preprints say leaderboard leads and test-time gains may be sampling luck. The fix is to audit, not to spend more.
- In a preprint, increasing reasoning effort in LLM shopping agents did not lead to statistically meaningful improvements in decision quality, but did lead to a statistically significant increase in reported confidence. Self-reported m:…1795
- In a preprint, only 13 of 173 models are statistically in the leading tier on SWE-bench Verified. Self-reported m:…1048
- In a preprint, in-distribution allocation gains in test-time budget allocation for neural combinatorial optimization solvers are often statistical artifacts rather than performance improvements. Self-reported m:…9521
- In a preprint, ScAn-Bench enables systematic evaluation of scaling law methodologies, which currently lack systematic evaluation across different model types. Self-reported m:…9509
More effort, same answer
In a preprint on LLM shopping agents, increasing reasoning effort does not lead to statistically meaningful improvements in decision quality, while increased reasoning effort leads to a statistically significant increase in reported confidence. The agents also exhibit poor calibration between reported confidence and actual correctness, and they show correlated error patterns, where multiple models tend to fail on the same tasks and select the same incorrect products. Longer thinking, in this setting, produces a more confident shopper, not a better one. m:…1795
The failure mode matters because it is invisible to the buyer. A model that has thought longer sounds more certain, and a team that pays for more reasoning tokens gets a more confident answer for the extra spend, but the extra spend buys certainty, not correctness. When several models fail the same way, the agreement between them looks like consensus and is actually a shared blind spot: the models are wrong together, and their agreement makes the wrong answer harder to question. m:…1795
The leaderboard may be measuring luck
A second preprint turns the same suspicion on the scoreboards. Public LLM leaderboards systematically overstate the performance of top-ranked models due to a 'winner's curse' where ranking rewards sampling luck. On the SWE-bench Verified benchmark, only 13 of 173 models are statistically in the leading tier. The paper's m-out-of-N bootstrap method reduces ranking bias, but the headline finding is that most of the leaderboard is not a ranking at all: it is a list of models whose luck has not yet run out. m:…1048
Read with the first paper, the pattern is one problem wearing two costumes. Inside a single model, extra effort inflates confidence without improving choices; across models, a leaderboard position inflates a reputation without proving a capability gap. In both cases the number that looks most decisive, the confidence score or the rank, is the number least tied to the underlying quality, and in both cases the remedy is the same: stop reading the headline number and check the uncertainty underneath it. m:…1795 m:…1048
How to check before you believe
A third preprint shows the same trap inside test-time budgeting. In-distribution allocation gains in neural combinatorial optimization solvers are often statistical artifacts rather than performance improvements, and non-uniform budget allocation guided by held-out sample statistics improves performance under distribution shift. The proposed correction procedure removes phantom gains while preserving real ones, which means the gains worth keeping are the ones that survive an audit, not the ones that look biggest in the first run. m:…9521
A fourth preprint offers the tooling to do that checking properly. ScAn-Bench-LLM and ScAn-Bench-VLM enable systematic evaluation of scaling law methodologies, because scaling law methodology currently lacks systematic evaluation across different model types. For developers, the practical advice follows from the four papers together: treat a rank-one lead as noise until a correction says otherwise, spend reasoning budget only where it changes decisions, and demand the audit trail before believing a gain. A team choosing between models should ask for the corrected tier, not the rank; a team buying reasoning tokens should test whether longer reasoning changes the decision, not whether it raises the confidence; and a team reporting a gain should run the correction procedure first and publish what survives it. m:…9509
What is still unresolved
- Does the confidence-without-correctness pattern hold outside shopping agents, in coding or reasoning tasks where verification is possible?
- How many published leaderboard leads would disappear if the winner's-curse correction were applied routinely?
- Will evaluation frameworks built on thousands of checkpoints become the standard way scaling claims are checked?
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.