Longer prompts make AI worse at reasoning — and the drop happens far before any memory limit is hit.

This explores why longer inputs make reasoning worse — and the corpus's most striking finding is that the damage happens *far below* the context window. On the FLenQA benchmark, reasoning accuracy fell from 92% to 68% with just 3,000 tokens of padding, and the drop was task-agnostic, uncorrelated with language-modeling quality, and survived chain-of-thought prompting Does reasoning ability actually degrade with longer inputs?. So the model isn't forgetting because it hit a memory ceiling. Something about the *presence* of more tokens — even irrelevant ones — degrades the reasoning itself.

A big part of the answer is that length hurts on both sides of the equation: not just longer inputs, but longer *thinking*. Accuracy on reasoning tokens follows an inverted U — push from ~1,100 to ~16,000 thinking tokens and benchmark scores can fall from 87% to 70% Does more thinking time always improve reasoning accuracy?. Optimal chain-of-thought length peaks in the middle and shrinks as models get more capable Why does chain of thought accuracy eventually decline with length?. The mechanism behind the mid-sequence collapse is telling: reasoning models 'wander' (explore invalid paths) and 'underthink' (abandon promising ones too early) — structural disorganization, not lack of compute Why do reasoning models abandon promising solution paths?. More tokens give the model more room to lose the thread.

There's also a deeper claim that length is a red herring for the real cause: *unfamiliarity*. Large reasoning models don't break at a complexity threshold — they break at novelty. They fit instance-level patterns rather than general algorithms, so a chain of any length succeeds if the model saw similar instances in training and fails otherwise, regardless of how long it is Do language models fail at reasoning due to complexity or novelty?. That reframes 'long inputs are hard' into 'long inputs contain more chances to hit something unfamiliar.' The same ceiling shows up when models are asked to sustain genuine backtracking: frontier systems score just 20–23% on constraint-satisfaction problems, revealing that fluent-sounding long reflection doesn't equal actual competence on unfamiliar structure Can reasoning models actually sustain long-chain reflection?.

There's a mathematical floor underneath all this, too. A Lipschitz-continuity analysis proves that longer reasoning chains *dampen* the propagation of input noise but can never eliminate it — a non-zero sensitivity floor exists structurally, so more context always carries some irreducible risk of throwing the answer off Can longer reasoning chains eliminate model sensitivity to input noise?. And length interacts badly with malformed inputs: reasoning models balloon into redundant chains on ill-posed or missing-premise questions because training taught them to produce reasoning steps but never taught them when to *stop* Why do reasoning models overthink ill-posed questions?.

The interesting upshot — the thing you might not have known you wanted to know — is that the fixes point away from 'add more tokens.' Verbosity turns out to be a single steerable direction in activation space (cut chains 67% with no accuracy loss) Can we steer reasoning toward brevity without retraining?; models can be trained to *route* between deep thinking and quick answers instead of always reasoning at length Can models learn when to think versus respond quickly?; and in multi-turn search, *capping* reasoning per turn preserves context for later evidence instead of burning it all upfront Does limiting reasoning per turn improve multi-turn search quality?. The consistent lesson: degradation with length isn't a capacity problem you solve with a bigger window — it's a discipline problem you solve by teaching models when to be brief.