tirtho149's picture
Add retriever study (code, predictions, one-pager), refresh response cache
874eec2 verified
|
Raw History Blame Contribute Delete
5.15 kB

Retriever pilot (branch retriever-pilot, 2026-11-17)

Question (Shreyan): can a small neural retrieval function, possibly with weights written by a bigger LLM, replace BM25 in MACE?

Setting: P1 MACE, Sonnet captions, seed 0, K = 4, the same per-query memory banks, 1,600 queries (1,000 IID, 600 held-out). Retrieval only, no diagnosis call. Proxies: hit@4 (true disease among the four retrieved entries; MACE gains only on hits), vote (majority label of the four), top-1 label. Equal-crop means (%), paired crop-stratified bootstrap for the hit@4 difference against BM25.

All retrievers score a (test caption, memory caption) pair from the same 16 features: whole-caption BM25 (the current retriever), per-field BM25 and Jaccard for the seven caption fields, and organ match.

Retriever hit@4 IID hit@4 held-out vote IID vote held-out top-1 IID top-1 held-out hit@4 vs BM25, IID [95% CI] held-out
BM25 (paper) 83.1 73.3 55.0 43.8 53.0 40.7
Trained linear 83.4 73.3 56.0 44.5 52.6 41.7 +0.3 [-1.5, 2.1] +0.0 [-2.7, 2.7]
Trained MLP (16-16-1) 82.0 73.0 55.0 43.8 51.5 40.5 -1.1 [-3.1, 0.8] -0.3 [-3.0, 2.3]
LLM weights, schema only 83.1 73.7 54.2 41.0 52.1 39.8 +0.0 [-2.2, 2.1] +0.4 [-3.0, 3.5]
LLM weights, schema + memory captions 82.5 73.2 54.3 41.2 51.6 41.7 -0.6 [-2.7, 1.5] -0.1 [-3.3, 3.0]
  • BM25 here reproduces the recorded P1 retrieval on 1,600/1,600 queries.
  • Trained models: leave-one-out pseudo-queries over the 20 memory banks only (74,376 pairs); no query image or label.
  • LLM weights: Sonnet writes 16 linear weights per crop (one call per crop and context, 20 calls, USD 0.35). Context A: feature definitions, crop, disease list. Context B: the same plus the crop's memory captions and labels.
  • Finding: no variant moves hit@4 by more than about one point; every interval includes zero. With lexical caption features, BM25 is already near what a small reweighting can reach, and giving the LLM the memory captions did not help.

Not tested: a network over learned text or image embeddings, LLM-written term weights, and diagnosis accuracy. Run python -m experiments.retriever_pilot.pilot (add --no-llm to skip the 20 calls; cached calls cost nothing).

Tuning (tune.py)

Tuned on a clean validation set, then one look at the 1,600 test queries. Validation: 2,215 training-fold images that are in the seed-0 K=16 memory but not in the K=4 memory MACE uses, and are never test queries, scored against that K=4 memory. Search: BM25 k1, b and organ repeat; DINOv2 image similarity added with weight alpha; at most cap entries per disease. 1,296 configs; objective hit@4, then true-disease entries among the four.

Retriever config (k1, b, organ x; alpha; cap) hit@4 IID hit@4 held-out true entries IID held-out vote IID vote held-out hit@4 vs BM25, IID [95% CI] held-out
BM25 (paper) (1.2, 0.75, 3); 0; 4 83.1 73.3 1.64 1.29 55.0 43.8
Image only (DINOv2) -; image; 4 81.0 69.3 1.55 1.11 50.7 36.0 -2.1 [-4.7, 0.5] -4.0 [-8.0, 0.0]
Text + image, no cap (2.0, 0.3, 0); 1.5; 4 86.3 74.2 1.75 1.25 60.1 40.7 +3.2 [1.3, 5.2] +0.8 [-2.0, 3.7]
Text + image, one per disease (2.0, 0.3, 0); 1.5; 1 89.3 80.3 0.89 0.80 58.3 42.7 +6.2 [4.2, 8.1] +7.0 [4.2, 9.8]
  • Adding image similarity to BM25 raises IID coverage (+3.2 hit@4, +5 vote) but not held-out (vote drops 3 points).
  • One entry per disease is the largest coverage win on both splits, but the VLM then sees each disease once, so the label majority signal that drives most of MACE's gain disappears. Whether coverage or repetition matters more can only be settled by a diagnosis run; these are retrieval proxies.
  • Retrieved ids for every selected config: out/tune_retrieved.json.

Diagnosis check (diag.py, 200-query efficiency panel, Sonnet, USD 4.32)

Same request as MACE except which four entries are listed. The BM25 arm went through the same code path and hit the cache of the recorded P1 calls on 200/200 queries with 200/200 identical predictions.

Method IID held-out vs MACE (BM25), IID [95% CI] held-out
SAGE 32.0 38.0 -23.0 [-32.0, -14.0] -22.0 [-30.0, -14.0]
MACE (BM25) 55.0 60.0
Text + image retrieval 56.0 63.0 +1.0 [-5.0, +7.0] +3.0 [-4.0, +10.0]
One entry per disease 47.0 58.0 -8.0 [-15.0, -1.0] -2.0 [-10.0, +6.0]
  • The retrieval-coverage win did not become a diagnosis win. One entry per disease raised hit@4 by 6 to 7 points but lowered IID accuracy by 8: the VLM needs the true label repeated, which matches the paper's finding that the retrieved labels carry most of the gain.
  • Text + image retrieval is directionally positive (+1 IID, +3 held-out) but far inside the noise on 200 queries. Telling a 1 to 3 point gain apart from zero would need the full 1,600 queries (about USD 25 for this one arm).