The Undetected Damage of Quantization on Retrieval — Top-1 score-gap certificate — What does it mean?
The news. On September 21, 2026, researchers at Sapienza University of Rome and collaborators posted The Undetected Damage of Quantization on Retrieval and How to Fix It. Across the models they tested, a quantized model that keeps its classification accuracy still changes 14–46% of its top-1 retrieval results. They trace the damage to one label-free number, the gap between the two highest scores, and use it for two fixes: spend extra bits on the layers that move the gap most (retrieval), and route low-gap inputs to full precision (classification). Read the paper →
Picture a race timed with a cheap stopwatch. For each runner, the stopwatch can be wrong by up to ε, and not by the same amount for everyone. The worst case for the leader is that their time reads ε too slow while the runner-up's reads ε too fast. So a lead is guaranteed to survive a bad stopwatch only when it is at least 2ε. Swap "runner" for "candidate" and "finishing position" for "score", and that is the paper's top-1 condition. Rounding weights to a coarse grid shifts every score the model computes, a class logit (the raw score for each class) in a classifier or a query-document cosine similarity in a retriever, by at most some ε. If the top score leads the second by at least 2ε, no shift of that size can change the winner. If it leads by less, a flip becomes possible, though not certain.
The two kinds of task run very different races. A classifier is trained with cross-entropy, which pushes the correct class far away from the others, so the typical input has a runaway leader. A retriever is trained with a contrastive loss that separates the matching document from sampled negatives, and nothing asks the best document to beat the second-best by a margin. So similar documents land at nearly the same coordinates and finish as a photo-finish pack. At 4 bits, the paper finds the contender set (everyone less than 2ε behind the leader) is empty for 74.3% of classification inputs on ViT-B/16 (a vision transformer) and for none of the retrieval queries. When a retrieval top-1 does change, the new winner is the old runner-up only 39.7% of the time; the rest come from deeper ranks, because the contender set often holds several candidates, not just the runner-up.
| Backbone (W4, RTN) | Top-1 changed as classifier | Top-1 changed as retriever | Source |
|---|---|---|---|
| ViT-B/16 | 6.2% | 45.5% | Table 1 |
| ViT-L/16 | 2.9% | 42.1% | Table 1 |
| Qwen3-Embedding-0.6B | 4.3% | 33.4% | Table 1 |
The averages hide this because they are built to tolerate small reorderings. On CLIP ViT-L/14 (an image-text model searching images by caption) at 4 bits, Recall@1 falls by only 1.3%, yet 10.9% of the queries whose correct image was ranked first lose it; 6.5% of all queries turn from right to wrong while 5.7% turn from wrong to right, and the two cancel in the average. On Qwen3-Embedding-8B, nDCG@10 drops 3.1% while 15.6% of queries lose their relevant top document. For a RAG pipeline, that is a different chunk in the prompt and a different generated answer, from a model the benchmark says is fine.
At a fixed ε, a separation ratio of at least 1 certifies that the winner cannot flip. Hold ε at 0.02 (illustrative) for one query, so the safety margin is 2ε = 0.04. A classifier's top logit leads the runner-up by 3.0 (illustrative): separation ratio 3.0 / 0.04 = 75, far above 1, so the answer cannot flip. A retriever's top cosine similarity is 0.71 and the second is 0.70 (illustrative): gap 0.01, separation ratio 0.01 / 0.04 = 0.25, and every document scoring above 0.67 is a contender. The paper's measured medians, at W4 with group size 128 pooled over image and text models, sit on the same two sides of 1: 5.73 for classification inputs and 0.081 for retrieval queries, and only 3.5% of retrieval queries reach 1.
Two fixes from one number
The share of at-risk inputs decides the fix. At W4 it is 94–100% of retrieval queries but only 9–26% of classification inputs. For retrieval, the fix has to be global: give the extra bits to the layers whose quantization moves the top-1/top-2 gap most. With a 3.5-bit average budget (half the layers at 4 bits, half at 3), ranking layers by this gap sensitivity recovers 60–73% of the benefit of a full extra bit, against 40–48% for ranking them by reconstruction error, the objective existing mixed-precision allocators use. Under RTN at that budget, the top-1 changes for 33.3% of queries instead of 43.1%. For classification, the exact certificate needs full-precision scores that are not available at run time, so the paper compares the quantized model's own gap against a threshold calibrated in advance on unlabeled data, and sends the low-gap inputs to the full-precision model; the paper reports this recovers most of the lost accuracy at a fraction of the full-precision cost.
The practical lesson is a new acceptance test. Before you ship a quantized embedder, compare its top-1 against the full-precision model per query, not its average metric. The paper also warns that the gap, not model size, predicts the damage: GTE-large, with about a twelfth of Qwen3-Embedding-4B's parameters, changes fewer top-1 results because its winners lead by more. Better quantizers such as GPTQ and AWQ reduce rounding error but, per the paper, do not remove the failure, because the pack is still a pack. The results cover weight-only quantization at 4 and 3 bits; activation quantization is untested.
Goes deeper in: LLM Internals → Quantization → What to Quantize