acceptodds
Withdrawn from ICLR 2027

Recovering Evidence Sufficiency Beyond Retrieval Similarity

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A retrieval system returns documents for every query, whether or not the answer is there. We would like to tell, from the embeddings alone, when the retrieved evidence is actually sufficient, before spending a generator on it. Prior work suggests this cannot be read off the geometry: fixed signals barely beat cosine similarity, so the geometry looks blind. We show the blindness was in the test, not the geometry. Those tests used an additive probe on the two vectors, which cannot represent any interaction between them and therefore cannot even represent cosine; it was asked to find a signal in a language that could not express it. Restore the interaction and the signal reappears, on all nine encoders we test and on a second, structurally different entailment corpus. Two facts explain the earlier null. The signal is not held in any few directions but spread redundantly across many, which is exactly why a fixed cosine readout misses it while a probe that pools directions recovers it. And the probe is nearly free: one dot product over embeddings the retriever already holds, orders of magnitude below a cross-encoder. That is cheap enough to gate every query before generation, with a finite-sample guarantee on how often sufficient evidence is wrongly rejected.

Primary area:
interpretability and explainable AI
Keywords:
retrieval-augmented generation, evidence sufficiency, embedding geometry, linear probing, concept erasure, conformal prediction, dense retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WsAAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims retrieval embeddings only look blind to evidence sufficiency because prior tests used additive probes. It shows an InferSent-style interaction probe beats cosine across 9 encoders, adds a cosine-matched swap test, an erasure analysis claiming the signal is “distributed,” and a conformal gate.

Critical strengths

None

Critical weaknesses
  1. The “blindness” being refuted is completely self-constructed. The abstract and intro attribute the null to “prior work,” but no cited work actually runs an additive (q,d) probe (+ global PCA) and concludes the geometry is blind: Joren et al. (2025) use LLM autoraters. Ethayarajh (2019) and Timkey & van Schijndel (2021) are about anisotropy, not sufficiency. The null in Section 3 is the authors’ own reproduction of an unnamed “prior practice.” Separately, the fact that a linear model on a concatenation cannot express interactions is exactly why InferSent introduced |u−v| in 2017. The entire premise of the paper seems self-contructed.

  2. Scoring 5 pairs with a cross-encoder costs nothing next to the LLM generation it precedes (orders of magnitude more per answer).

Questions for authors
  1. Look at the weaknesses section.
  2. Why is theorem 1 even stated if there is no use for it?
1 comment
AuthorsOpenReview ↗

Response to Reviewer WsAA

Thank you. Your first weakness is correct, and the controls prompted by the other reviews show that the positive claim built on that premise does not hold either.

W1. No cited work ran an additive probe and concluded the geometry is blind; the null was ours. Moreover, in our own data the additive probe on [q; d] matches the interaction probe on ArguAna (0.704 against 0.685), and on SciFact the interaction probe's gain disappears once gold documents are held out (general response, point 1). The premise and the claim are both withdrawn.

W2. Agreed. With five retrieved documents, a cross-encoder on five pairs is negligible next to the generation it precedes. The cost argument is withdrawn.

Q2. Theorem 1 supported no empirical claim and should not have been stated as it was.

Reviewer_tyvoOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether a dense retriever's embeddings alone can indicate if the retrieved evidence is sufficient to answer a query, before running a generator. The authors argue that prior negative findings come from additive probes on [q; d], which cannot represent query–document interactions. They show that a logistic probe, [q; d; |q−d|; q⊙d] outperforms cosine similarity across nine bge/e5/gte encoders on SciFact and ArguAna. Supporting evidence comes from Climate-FEVER and from a cosine-matched swap test intended to show that the probe reads content rather than similarity. The paper then analyzes the learned weights and uses linear concept erasure to argue that the signal is distributed across many directions. Finally, it turns the probe into a split-conformal gate and compares its cost against a trained cross-encoder.

Critical strengths
  • Deciding retrieval sufficiency from existing embeddings before generation is a relevant problem for RAG. The interaction probe costs almost nothing on top of retrieval and improves over cosine on all nine encoders on SciFact and ArguAna, under a clean nested cross-validation setup.
  • Reusable audit protocol and transparent reporting of failure cases. The cosine-matched swap test with an explicit similarity certificate is a useful idea for checking whether a sufficiency signal reads content rather than similarity. The paper also openly reports where things break, including the conformal gate failing under task shift (Table 6) and no gain on FiQA.
Critical weaknesses

W1. The main comparison does not isolate the claimed effect

  • All headline results compare a trained probe against untrained cosine. Because the interaction probe contains cosine as a special case (uniform w), beating cosine is expected and does not by itself support the claim that interaction terms are what matter.
  • The key control, a trained additive probe on [q; d], is only described in words ("tracks cosine") and never reported numerically. Other cheap trained baselines are also missing, e.g., a non-linear probe on [q; d] and standard post-retrieval QPP features.

W2. The mechanistic explanation does not explain why cosine fails

  • Section 6 only shows that the signal survives erasure of a small subspace. A redundant signal with consistent sign would be captured well by cosine, which sums all directions uniformly.
  • The actual reason cosine fails appears to be the sign reversal in Section 5 (negative-mass readout on ArguAna), yet the Section 6 title and the Introduction present distributed encoding as the cause.

W3. The swap test is weak evidence for content reading

  • Probe AUROC is only 0.54–0.59, with CI lower bounds close to 0.5.
  • Cosine AUROC on the matched pairs is above 0.5 for all nine encoders. Since the probe includes a cosine-like term, part of its margin may come from residual similarity.
  • The "pre-registered" rule has no reference, and it is unclear whether swap-test queries were excluded from probe training.

W4. Overclaiming and internal inconsistencies

  • The rhetoric exceeds the evidence ("the diagnosis is unambiguous", "it is the cause"). Theorem 1 is a standard upper bound that does not support any empirical claim.
  • Figure 1 reports nine-encoder mean Δ while plotting only three encoders.
  • Section 5 reports results on cosine-matched ArguAna pairs, while Section 4 states that such pairs essentially do not exist.
  • "Coherent irrelevance" is referenced in Section 6 but never defined.
Questions for authors

Please see the weakness above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer tyvo

Thank you. W1 and W3 pointed at the two controls that decided the matter.

W1. Table R1 now reports the trained additive probe that the submission described only in words. It does not track cosine: on SciFact it is below cosine (0.713 against 0.801 with the submitted folds, 0.567 held out), and on ArguAna it matches the full probe (0.704 against 0.685). So the interaction terms are not what matters, which is your point, and the contrast the paper was built on does not exist in our data. We did not run post-retrieval QPP features.

W2. On ArguAna a free diagonal does beat a non-negative one (0.652 against 0.595), so negative weights help a diagonal readout, as you say. But a probe on the document alone exceeds both (0.721). The mechanism is not a property of the pair, and the section is withdrawn.

W3. With query and gold document held out, the full probe reaches 0.570 [0.546, 0.595], 0.553 [0.531, 0.577] and 0.564 [0.540, 0.589] on SciFact, and never exceeds the document-only probe (Table R2). The swap queries were not held out for Climate-FEVER; that was an error. The word "pre-registered" had no public registration behind it and should not have been used.

W4. All four points are correct. On the inconsistency between Sections 4 and 5: Section 4 was the wrong one. Cosine-matched ArguAna pairs exist (920 to 1,172 per encoder); our statement that they do not was wrong. Theorem 1 supported no empirical claim. Figure 1 mixed three plotted encoders with nine-encoder means. "Coherent irrelevance" was undefined.

Reviewer_KQrEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authos ask whether retrieval embeddings alone can indicate if retrieved evidence is sufficient to answer a query, before a generator is run. It argues that prior “blindness” results came from additive linear probes on [q;d][q; d], which cannot represent any q–d interaction (not even cosine). Using the standard InferSent feature set [q;d;∣q−d∣;q⊙d][q; d; |q−d|; q\odot d], a logistic probe beats cosine on 9/9 BGE/E5/GTE encoders on SciFact (AUROC), on ArguAna (where cosine is at chance), and on Climate-FEVER, with no gain on FiQA. A cosine-matched swap test certifies that part of the signal is content rather than similarity (AUROC 0.54–0.59 on SciFact, 0.83–0.89 on Climate-FEVER). Weight-space analysis on three small encoders shows the learned readout is stable, roughly cosine-like on SciFact, and runs against the identity direction on ArguAna. A LEACE-based erasure sweep shows the signal survives removal of up to 32 dimensions of the probe’s own subspace, which the authors interpret as a distributed, redundant encoding and offer as the mechanism behind cosine’s blindness. Finally, a split-conformal gate built on the probe escalates fewer queries than a cosine gate at fixed false-negative rate, and the guarantee is shown to break asymmetrically under task shift.

Critical strengths
  1. The cosine-matched swap test is well-designed. Reporting the rank-displacement condition as a control that cosine must win, and attaching a certificate that cosine AUC lies in [0.45, 0.55], is a reusable discipline that other sufficiency papers lack. The result is honestly reported as modest.

  2. The conformal gating section does something most “cheap signal” papers skip: it measures tail behavior rather than AUROC, and it shows that a threshold calibrated where cosine works (SciFact) is catastrophic where it does not (ArguAna, FNR 0.88–1.00). This is a practically important negative result.

Critical weaknesses
  1. The authors putThe central frame on a not clearly attributed claim, and the sufficiency signal that survives the paper’s own examination is small. The paper reproduces a “null” with an additive probe “as in prior practice”, but neither cited source did this: Joren et al. build an LLM-based sufficiency classifier and never claim geometry is blind, and Ethayarajh / Timkey & van Schijndel study anisotropy, not sufficiency. So the “hypothesis-class artifact” diagnosis is correcting a claim the paper itself constructed. More importantly, on the one dataset where the full audit is run across all encoders (SciFact), the content signal at fixed similarity is AUROC 0.54–0.59, and the paper’s own Section 5 analysis says the SciFact readout is essentially a recalibrated cosine (alignment with the uniform direction +0.54 to +0.68, little negative mass). That means most of the +0.083 headline is likely a learned per-coordinate reweighting of cosine, which is anisotropy correction, not a sufficiency signal. The large gains are on ArguAna, which the authors concede is not a sufficiency task. The paper needs a whitened or learned-diagonal cosine baseline on SciFact and Climate-FEVER to separate these two explanations, and it needs to explain why Climate-FEVER’s swap AUROC (0.83–0.89) is so far above SciFact’s (0.54–0.59). A plausible alternative is that cosine-matched distractors drawn from a 5.4M-sentence Wikipedia corpus are topically unrelated to the claim, making the test much easier there, not that the signal is “stronger”.

  2. The “mechanistic explanation” in section 6 does not actually explain cosine’s blindness, and it is overclaimed. The authors claim tthe signal is distributed across many directions, therefore a fixed isotropic readout (cosine) cannot see it. This does not follow. Cosine on normalized embeddings is the fixed bilinear form q⊤Idq^\top I d ( w=1w = 1). It fails to pick up a signal when the discriminating weight vector w⋆w^\star is poorly aligned with the all-ones direction. Whether w⋆w^\star is concentrated on one coordinate or spread over hundreds has clear meaning on that alignment: a signal living entirely in a single coordinate direction orthogonal to 1/d1/\sqrt{d} is just as invisible to cosine as a diffuse one. So “distributed” is neither necessary nor sufficient for cosine blindness.

Questions for authors
  1. Which document d is paired with each query when training and evaluating the probe: the top-1 retrieved document, each of the top-5, or something else? And how many training pairs per dataset (SciFact has ~1.4k claims)? If d is the top-1 and the label is “gold anywhere in top-5”, please say how often d itself is the gold document, since this affects what the probe is learning.

Q2. Please report a whitened cosine (or learned non-negative diagonal q⊤diag(w)dq^\top \text{diag}(w) d with w≥0w \ge 0) baseline on SciFact and Climate-FEVER. How much of the +0.083 and +0.16/0.20 gains survive this baseline?

Q3. Why is the swap-test AUROC on Climate-FEVER (0.83–0.89) so much higher than on SciFact (0.54–0.59)? Can you characterize the matched distractors (topical overlap with the claim, whether they are from the same article) to rule out the corpus-size explanation above?

Q4. Which prior work ran an additive probe on [q; d] and concluded geometry is blind? If none, please reframe the contribution accordingly.

Q5. Please give the logical link from “distributed encoding” to “cosine cannot express the signal”, or drop the causal language. Is the explanation in section 5 (negative mass, misalignment with the uniform direction) not already sufficient?

1 comment
AuthorsOpenReview ↗

Response to Reviewer KQrE

Thank you. Your W1 was right, and the effect is weaker than you suspected.

W1. With gold documents held out, the SciFact gain is not reduced to a reweighted cosine: it is gone. The full probe is below cosine by 0.02 to 0.07 on all three encoders, and the non-negative diagonal is within ±0.03 of cosine (general response, point 1). The frame was also wrong: no prior work ran the additive probe and concluded the geometry is blind.

W2. Accepted. Beyond your argument, the document-only control shows that the ArguAna readout analysed in Sections 5 and 6 is not a pair-level readout at all. Those sections are withdrawn.

Q1. Each query is paired with each of its top-5 retrieved documents, and the label is whether that document is gold. It is a pair-level label; Section 2 described it at query level, which was inaccurate. SciFact: 5,545 pairs from 1,109 queries, 930 to 986 positive depending on the encoder; the top-1 document is gold for 58–62% of queries. Climate-FEVER: 7,675 pairs from 1,535 queries, 794 to 1,019 positive; top-1 gold for 20–27%. ArguAna: 7,030 pairs from 1,406 queries, 718 to 969 positive; top-1 gold for 0.1–0.2%.

Q2. Table R1, column "diag, w ≥ 0".

Q3. The gap was our evaluation error together with a document prior. The Climate-FEVER probe was fitted on all pairs, including about half of the swap pairs themselves. Held out, the full probe gives 0.593, 0.713 and 0.573, and a document-only probe gives 0.641, 0.727 and 0.640. On your alternative explanation: the mean cosine between gold document and matched distractor is similar on the two corpora (SciFact 0.77, 0.87, 0.85; Climate-FEVER 0.71, 0.84, 0.86), so we find no evidence that the Climate-FEVER distractors are farther from the gold. The difference came from the leak.

Q4. None.

Q5. The step does not hold and the claim is withdrawn.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.