The convenient explanation that circulated through the model community for two years went like this: longer context, dumber model. It had hardened into a consensus. A 2025 Chroma technical report named the degradation "context rot" across 18 frontier models, and an earlier paper, "Context Length Alone Hurts LLM Performance Despite Perfect Retrieval" by Du et al., offered a clean and convincing proof: mask every irrelevant token out of attention while keeping position indices, and models still degrade by 13.9% to 85% on math, question answering, and coding. If hiding the distraction entirely still costs you, length itself must be the culprit.
On 3 November 2026, Yinuo Liu of UC San Diego posted a paper aimed squarely at that conclusion (arXiv:2611.04169). The design rests on one principle: every comparison holds length and evidence position fixed and changes only what surrounds the evidence. Five open models, 0.5B to 3B parameters, four model families. The task is key-value retrieval: the context hides one statement, "The secret code of the river is 4821.", and the closing query asks for the river's secret code, scored as the log-probability of the correct value under teacher forcing. Four conditions: nothing filled in, neutral filler (random numbers, random words, one repeated sentence), competing statements in the same template with keys drawn from 51 common nouns, and filler masked out of attention with positions preserved.
Length is nearly free, competition is not
On Qwen2.5-1.5B, leaving neutral filler visible keeps the median loss at or below 0.006 nats at every length from 8K to 32K, with accuracy between 0.96 and 1.00. Swapping in competing statements at the same lengths pushes the median loss to 2.82, 3.94, 4.03, and 4.76 nats, and accuracy down to 0.11 to 0.21. What breaks retrieval is never the tens of thousands of tokens themselves. It is whether anything inside them competes with the evidence for attention.
The sharper control separates conflict from format. The paper isolates statements that share the queried key but assign it a different value: across five models, contexts with no same-key conflict lose 0.02 to 0.34 nats on average, while contexts with at least one lose 1.38 to 3.24 nats. Accuracy is 0.97 to 1.00 on one side and 0.25 to 0.52 on the other. At a fixed length of 2048 tokens, raising same-key conflicts from 0 to 6 or more sends Qwen2.5-1.5B along a climb of 0.03, 1.37, 2.24, 2.48, 2.71, 3.26, and 3.45 nats. In the other direction, distractors that share only the format but not the key explain at most 0.3% of the variance in loss.
There is one accident the author flags in the repository himself. When generating competing filler, keys were drawn from 51 words without excluding the queried one, so roughly one filler sentence in 51 restates the queried key with a different value, and the context contradicts itself. He puts it plainly: this was an accident. The repository contains a measure_v4.py that fixes the problem and adds a controlled single-rival arm, but it has not been run, and no v4 result appears anywhere in the repository.
How much attention is stolen decides how much accuracy is lost
The paper then feeds "how much attention is left on the evidence" into a regression. It selects eight layer-head pairs by their attention mass on the evidence within the short context, and reads the attention mass a* that the final query token sends to the evidence tokens. Across 126,473 trials, log(1/a*) explains 35% to 58% of the variance in loss, while log length explains at most 2% to 8%. Per model, Qwen2.5-1.5B gives 0.58 against 0.05, OLMo-2-1B gives 0.35 against 0.02, and Llama-3.2-3B gives 0.45 against 0.08.
Two more details are worth keeping. Attention on the evidence falls from 0.51 to 0.77 in the short context to 0.17 to 0.31 under competing filler. And the relation has a knee: on Qwen2.5-1.5B, the loss stays near zero until a* falls to about 0.35, then climbs steeply past that point. The model tolerates losing about half of the attention it gives the evidence, and then falls off a cliff.
Masking is not a clean control for length
The sharpest move in this design is to go back and audit the evidence behind the "hiding the distraction still hurts" claim. The paper replicates that masked protocol on MMLU: 1,000 questions with 3,750 to 30,000 masked tokens inserted, where Llama-3.2-3B falls from 51.7% to 40.5% and Qwen2.5-7B falls from 65.8% to 48.1%. The replication holds, so this is not specific to one model family.
But the same masking dropped into the retrieval task reverses the conclusion. At 32K tokens, masked neutral filler costs 46 accuracy points, dropping from 0.98 to 0.52, while the identical filler left visible costs nothing at all. Averaged over the three neutral fillers at 32K, accuracy is 0.96 to 0.99 when visible and 0.48 to 0.54 when masked. The paper's explanation is that masking leaves long positional gaps the model never encounters in training, which are not the same thing as irrelevant tokens, so what it measures is not the cost of length.
There is a counterexample on MMLU as well. For Qwen2.5-1.5B at 4,096 tokens, accuracy is 59.1% with no filler, 57.5% with visible neutral filler, and 53.9% with masked filler. But with whitespace padding instead, it drops straight to 36.0%. Length is not entirely innocent. The wound it leaves on reasoning tasks is simply not the same wound it leaves on retrieval.
So what
If the result holds, engineering priorities invert. The real budget of a context is not its token count but the number of statements competing with the evidence. For retrieval-augmented generation, agent memory, and long-document pipelines, deduplicating and resolving conflicting statements before they reach the model is worth more than truncating the context. And because a* needs only a single forward pass, it can serve as a pre-deployment health check: once attention on the target falls past a threshold, resolve the conflict first rather than buying a bigger window.
The paper is also explicit about its own limits. Every result comes from a single synthetic retrieval task with models of at most 3B. Because a* and the loss are measured in the same forward pass with no causal intervention, the relation is predictive rather than causal. The conflict analysis was not preregistered. Every context of 4,096 tokens or more contains at least one conflict, so the branch covering format-only competitors above 2K was never measured. And the collision every 51 filler sentences that the author also calls an accident sits unrepaired, with the real controlled experiment still unrun in the repository.
So the accurate statement is not that length is harmless. It is that in retrieval the bill for length has to be recomputed, and competition is the bill that actually comes due. Whoever gets that arithmetic right can explain why the same advertised context window works for one model and falls apart the moment another touches it.