{"id":"a0e36ef8-a793-4d6f-a13f-7aa272846e62","arxiv_id":"2512.00919","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Augmented Spectral Feature Learning biases the learned treatment–instrument features toward the outcome, recovering NPIV structural functions that spectrum-only SpecIV features miss.","lead":"By adding an outcome-aware term to spectral feature learning for instrumental variable regression, this paper makes learned causal features useful even when the causal function is poorly captured by the dominant treatment–instrument modes. The method is tested on synthetic, dSprites, and off-policy evaluation benchmarks, where it often beats the outcome-agnostic baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bad-scenario guarantee (Eq. 18) rests on a false inequality: the claimed γ1 lower bound is equivalent to √(1+x)>√x+1.","rationale":"The most load-bearing point is not the oracle selection (though real), but an algebraic error inside the proof of the central guarantee. Section 4's bad-scenario derivation claims γ1(δ)>λ_kδ‖s1‖-(1-λ_k). Starting from the paper's own Eq. (8) and its expression γ1=λ_k√(1+δ^2‖s1‖^2)-1 (valid because ∥Λ_Nbar∥=λ1=1 for a conditional-expectation operator), this inequality rearranges to √(1+x)>√x+1, which is false for all x>0. Consequently the condition 7‖q1‖/‖s1‖+3E_1λ_k≤κλ_k^2 and the advertised 'one feature' conclusion are unsupported. The reader's oracle concern is legitimate and adds to this, but the proof error is the primary blocker: even with perfect knowledge of k and δ, the guarantee does not follow as written. An independent re-derivation or numerical check settles it immediately. If the algebra is corrected, the verdict could return to CONDITIONAL pending data-driven δ/k selection; as written, the central theoretical claim should be rejected.","tokens_in":30831,"tokens_out":14301,"duration_ms":120037,"concrete_test":"Re-derive Eq. (18) from Theorem 4: for N={k}, λ1=1, write γ1(δ)=λ_k√(1+x)-1 with x=δ^2‖s1‖^2. Substitute the paper's δ and check whether γ1 > 6(1-λ_k). The inequality reduces algebraically to √(1+x) > √x+1, which is false (e.g., x=1). If the authors instead provide a corrected lower bound, re-verify the condition 7‖q1‖/‖s1‖+3E_1λ_k≤κλ_k^2 and the displayed error bound for λ_k∈{0.1,0.5,0.9}; this will show whether the 'one feature' claim can be salvaged.","verdict_should_be":"REJECT","load_bearing_attack":"The advertised bad-scenario recovery guarantee (Eq. 18) is not established by the proof as written. In the bad scenario, N={k}; because T is a conditional-expectation operator, λ1=1 and ∥Λ_Nbar∥=1, so Eq. (8) gives γ1(δ)=λ_k√(1+δ^2‖s1‖^2)-1. The paper states γ1(δ)>λ_kδ‖s1‖-(1-λ_k) after setting δ=7(1-λ_k)/(λ_k‖s1‖). Subtracting 1 and dividing by λ_k, this is equivalent to √(1+x) > √x+1, which is false for every x=δ^2‖s1‖^2>0. Thus the displayed lower bound 6(1-λ_k), the condition 7‖q1‖/‖s1‖+3E_1λ_k≤κλ_k^2, and the headline conclusion 'one feature suffices whenever ‖s1‖/‖q1‖≫λ_k^{-2}' do not follow from the given derivation. The result may be repairable with a different δ choice, but the current proof leaves the central theoretical claim unsupported. The additional oracle requirements on k and δ noted by the reader are real but secondary; even granting them, the proof fails.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Augmented Spectral Feature Learning, an outcome-aware variant of SpecIV for nonparametric instrumental variable regression. The method replaces the target-agnostic spectral loss of SpecIV with a contrastive loss derived from an augmented operator T_δ = [T | δ r_0]. Proposition 1 establishes that minimizing this loss is equivalent to best rank-d approximation of T_δ. Theorem 1 gives a high-probability 2SLS excess-risk bound under sub-Gaussian assumptions, and Section 4/Appendix B.3 develop perturbation bounds intended to show that in a 'bad scenario' of spectral misalignment the method recovers the structural function with a single feature whenever the signal-to-noise ratio dominates the spectral decay. Experiments on synthetic data, dSprites benchmarks, and off-policy evaluation show gains over standard SpecIV for suitable δ.","tokens_in":31268,"tokens_out":11175,"duration_ms":109548,"significance":"The core idea is well motivated: making spectral feature learning outcome-aware directly targets the known failure mode of SpecIV when h_0 is misaligned with the top singular functions of T. The paper has several strengths: the data-splitting design (independent feature-learning and 2SLS datasets) avoids circularity in the statistical analysis; Proposition 1 is clean; Theorem 1 is a useful high-probability 2SLS bound; and the experimental section is substantial, including a genuinely harder dSprites task and an OPE setting. The authors are also candid about limitations, explicitly flagging the unquantified optimality gap E_d and the lack of theoretical justification for stage-2-loss model selection. However, the central 'bad scenario' guarantee — the paper's main advertised advantage over SpecIV — is not established by the current proof because of an invalid inequality in Appendix B.3. The idea may be repairable, but the theoretical claim as written is unsupported.","major_comments":[{"comment":"The one-feature recovery guarantee rests on an invalid inequality. From Eq. (8) with N={k}, Λ_d=λ_k, the correct gap is γ1(δ)=λ_k(√(1+δ²‖s1‖²)−1), not λ_k√(1+δ²‖s1‖²)−1 as written. With the text's formula, the claimed inequality γ1(δ)>λ_kδ‖s1‖−(1−λ_k) is equivalent, after dividing by λ_k, to √(1+x)>√x+1 for x=(δ‖s1‖)², which is false for every x>0. With the correct Eq. (8) formula, the claimed lower bound γ1>6(1−λ_k) requires conditions on λ_k that are not stated or proved. Therefore Eq. (18), the condition 7‖q1‖/‖s1‖+3E_1λ_k≤κλ_k², and the headline conclusion 'one feature suffices whenever ‖s1‖/‖q1‖≫λ_k⁻²' do not follow from the given derivation. The choice of δ (or the statement of the required spectral gap) must be corrected.","section":"Appendix B.3, 'Bad scenario' paragraph; Section 4, Eq. (18)"},{"comment":"The advertised recovery guarantee is oracle-dependent in a way that is not acknowledged. The proof assumes the practitioner knows the singular index k on which h_0 concentrates and sets δ=7(1−λ_k)/(λ_k‖s1‖) using the true signal norm ‖s1‖. No data-driven procedure for identifying k is given, and Section 6.4's δ-selection rules are heuristic; Section D.8 explicitly concedes that stage-2-loss model selection has no theoretical justification. As a result, the theorem is a conditional oracle bound rather than a guarantee for the proposed algorithm. The paper should either provide a practical selection rule or clearly state that the theoretical result assumes oracle knowledge.","section":"Section 4, bad-scenario paragraph"},{"comment":"The end-to-end guarantee is conditional on the optimality gap E_d(θ,ω,δ)=‖T_δ^(d)−Ψ_θ[Φ_θ^*|ω]‖, which the paper explicitly leaves unquantified ('requires architecture-specific generalization bounds... an open problem'). Consequently the corollaries in the good and bad scenarios contain an additive term that could be arbitrarily large, and the results do not establish that training the neural networks with the contrastive loss attains the advertised rates. This is a known and acknowledged limitation, but it should be more prominently qualified in the abstract and conclusions, since the stated 'full generalization error bound' overstates what is proved.","section":"Section 4, Eq. (7) and Theorem 4"}],"minor_comments":[{"comment":"The notation 'Z⊥⊥/X' is a typo; it should express conditional independence, e.g., Z⊥⊥X with a slash through the independence symbol.","section":"Section 2, Eq. (2)"},{"comment":"The 'if and only if' statement assumes uniqueness of the best rank-d approximation of T_δ. If the singular values of T_δ have ties, the minimizer is not unique; the statement should say 'one of the minimizers' or add an explicit spectral-gap assumption for T_δ.","section":"Proposition 1"},{"comment":"The displayed coefficient (5‖h0‖+λ_k²‖s1‖)/(λ_k²‖s1‖) appears to be a typo; presumably it should be 5‖h0‖/(λ_k²‖s1‖)+1. The condition '7‖q1‖/‖s1‖+3E_1λ_k≤κλ_k²' also does not match the substitution into Theorem 4's condition, even before accounting for the invalid γ1 bound.","section":"Eq. (18)"},{"comment":"The stage-2-loss model-selection strategy is used in the OPE experiments, despite the paper's own Section D.8 explaining that its theoretical justification 'remains elusive.' The main text should mark this strategy as heuristic, since a reader may otherwise take it as validated.","section":"Section 6.4 / Section D.8"},{"comment":"The sentence 'Since δ was automatically tuned ... and took the values of 1/10−3/10−2 for Cartpole/Mountain Car/Catch' is hard to parse; it should read something like 'took the values 1, 10⁻³, and 10⁻² for Cartpole, Mountain Car, and Catch, respectively.'","section":"Section 6.3 / Section D.7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a genuine contribution and the empirical results are encouraging, but the central theoretical claim in the bad scenario is currently unsupported due to the inequality error in Appendix B.3. This is a load-bearing issue; I would ask the authors to repair the proof or substantially weaken the claimed guarantee before publication. The oracle-dependence of δ and k should also be addressed in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the core construction here is real and worth knowing, but the paper's headline theoretical claim—that in the bad spectral misalignment case one feature suffices—is not proven. The proof of the key inequality in the bad scenario (Section B.3, around Eq. 18) relies on γ1(δ) > λ_k δ‖s1‖ − (1−λ_k). Substituting the chosen δ and the actual γ1 turns this into √(1+x) > √x + 1, which is false for every x>0. So the displayed lower bound 6(1−λ_k), the condition 7‖q1‖/‖s1‖ + 3E_1λ_k ≤ κλ_k^2, and the 'one feature recovers h0' conclusion do not follow from the derivation as written. This is not a minor typo; it's the advertised advantage over SpecIV. The result may be repairable with a different δ, but the current manuscript doesn't do it.\n\nWhat's genuinely good: the outcome-augmented operator Tδ = [T | δ r0] is a simple, elegant step. Proposition 1 makes the equivalence to a low-rank approximation precise, and the contrastive loss in Eq. (5) is a practical way to optimize it. The general 2SLS bound (Theorem 1) is standard but cleanly proven in high probability. The experiments are careful: the new dSprites structural function is motivated, the synthetic setup controls the spectrum, and the OPE application shows the method can help when the outcome changes iteratively. Small positive δ consistently helps, and the authors are candid about what they don't know—they explicitly leave the optimality gap open in Appendix B.3 and admit in D.8 that 2SLS-loss model selection is unprincipled.\n\nSoft spots, in proportion: the false inequality is the big one. The reader's complaint about oracle knowledge of k and δ is secondary but real: the bad-scenario guarantee assumes you know which singular index h0 concentrates on and sets δ from the true signal norm. No data-driven procedure is given. And there is no code/data release, which hurts reproducibility. None of these would be fatal if the theory were sound; together they mean the paper needs significant revision.\n\nBottom line: this is a paper for the spectral-NPIV community. The idea is worth a serious referee—the construction and experiments merit attention—but the current proof does not support the central claim. I'd send it to review with the expectation that the authors fix the bad-scenario analysis, and I'd want to see code/data before accepting.","headline":"New outcome-augmented spectral loss is a good idea and the experiments are encouraging, but the main 'one feature suffices' theorem rests on a false inequality and needs repair before the central claim can be trusted.","tokens_in":31693,"tokens_out":5777,"would_cite":true,"duration_ms":47809,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that making spectral features outcome-aware — by adding a column δr₀ to the treatment–instrument operator — turns the worst spectral misalignment case into a one-feature recovery problem.","keywords":["nonparametric instrumental variable regression","spectral feature learning","outcome-aware features","augmented operator","contrastive loss","spectral misalignment","two-stage least squares","off-policy evaluation"],"falsifier":"Take a finite-rank IV problem where the structural function equals one deep singular function exactly, compute the top singular subspace of Tδ with δ given by the paper's formula, and check whether the singular-value gap is positive and the top left singular vector aligns with the intended direction; then train the one-feature model and compare the L2 error to the bound. If the error does not fall as the signal-to-noise ratio grows with the singular value fixed, the mechanism fails.","tokens_in":30759,"feed_emoji":"🎯","tokens_out":7829,"duration_ms":75164,"temperature":0.7,"pith_summary":"The paper takes on a known failure of spectral feature learning in nonparametric instrumental variable (IV) regression: features are learned from the operator linking treatments to instruments, so they can be useless when the true causal function is concentrated in the tail of that operator's spectrum. Its central move is to augment the operator with a column aligned with the outcome, defining an augmented operator Tδ and a new contrastive loss whose global minimiser is exactly the rank-d truncated SVD of Tδ. The paper proves a non-asymptotic, high-probability error bound for the resulting two-stage least squares estimator, and in the 'bad' scenario — where the structural function is aligned with a deep singular function — shows that one outcome-aware feature can recover the structural function where the old method would need many features. Empirically, a small positive augmentation strength helps even in well-aligned settings, and on a deliberately misaligned image-based benchmark the outcome-aware version matches or beats a competitive deep-feature IV baseline. A reader should care because spectral methods for IV are useful but fragile, and this is a direct fix with a proof of the mechanism.","feed_headline":"One feature recovers the causal function in misaligned IV regression","feed_subtitle":"Outcome-aware spectral features keep IV estimates accurate when the signal hides in the spectrum's tail.","key_machinery":"The load-bearing object is the augmented operator Tδ = [T | δr₀], which maps (h, a) to Th + aδr₀; adding the outcome column δr₀ bends the top singular subspaces toward components of the structural function that the original operator T ignores. The paper proves that the new contrastive loss Lδ(θ,ω) equals ‖Tδ − Ψθ[Φθ* | ω]‖_HS² − ‖Tδ‖_HS², so minimising it is exactly truncated SVD of Tδ. A singular-value gap γδ controls the perturbation between the subspaces of T and Tδ, and that gap is what makes one-feature recovery possible in the misaligned regime.","core_discovery":"On its own terms, the paper's claim is that the outcome-agnostic failure of spectral IV regression is not intrinsic to spectral features, but only to the operator being decomposed. If the operator is changed from T to Tδ = [T | δr₀], then minimising the augmented loss Lδ(θ,ω) = ‖Tδ − Ψθ[Φθ* | ω]‖_HS² − ‖Tδ‖_HS² is equivalent to computing the rank-d truncated SVD of Tδ. When the structural function h₀ concentrates on a singular function deep in the spectrum of T, the paper proves that for large enough δ the top singular subspace of Tδ aligns with that signal, and the estimation error becomes approximately (1/λ_k²)(‖q₁‖/‖s₁‖) plus a representation-learning error; thus whenever the signal-to-no","pith_inferences":["The rank-one augmentation is a template beyond IV: any spectral decomposition of a conditional operator can be steered by appending a rank-one column proportional to a downstream target, provided the contrastive loss admits the same Hilbert–Schmidt identity. This suggests analogues in state-space model learning and in solving conditional moment restrictions more broadly.","The bad-scenario guarantee is oracle-dependent: choosing the right spectral index and augmentation strength needs the true signal norm. A data-driven rule that estimates these from the fitted operator would turn the theory into a practical procedure; the paper does not provide one.","The higher-rank extension with moments such as E[Y^k|Z] is only sketched, but if the noise is heteroscedastic or the conditional distribution is asymmetric, those extra moments carry information the first moment misses. A full perturbation analysis for rank-K augmentation could make the method robust beyond mean-zero confounders.","The consistent benefit of small positive augmentation even in well-aligned settings hints that the augmentation acts as a regulariser, biasing features toward predictive directions at negligible cost to the operator approximation; testing this interpretation across a wider range of signal-to-noise ratios would be a cheap, direct extension."],"forward_implications":["Spectral IV estimators no longer need to learn many features when the signal sits deep in the spectrum; one outcome-aware feature can be enough, reducing feature dimension and statistical cost in misaligned regimes.","Because the loss identity holds for any parameterisation, existing contrastive feature-learning pipelines can be made outcome-aware by adding the regularisation term, without changing the downstream 2SLS step.","The augmented loss gives a principled way to measure how much of the structural function lies in the learned feature span, which can be used to compare feature sets and to guide the augmentation strength.","In off-policy evaluation, where the outcome changes at every iteration of the value iteration, an outcome-agnostic spectral method can lose the relevant direction; the augmented method tracks the shifting signal and remains competitive.","The generalisation bound is high-probability rather than in expectation, so practitioners get a finite-sample guarantee that degrades gracefully with the ill-posedness of the learned representation."],"fun_headline_variants":["Outcome-aware features fix spectral IV regression misalignment","New loss makes spectral IV learn the right signal","Augmented operator aligns spectral features with the outcome","When spectral features miss, outcome-aware learning recovers","Task-specific spectral features: the cure for misaligned IV"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The misaligned-regime recovery guarantee assumes the user already knows the spectral index where the signal sits and sets the augmentation strength from the true signal norm, and that the feature networks reach a small optimality gap in training; neither is guaranteed by the theory.","fun_headline_variants_meta":{"raw":{"variants":["Outcome-aware features fix spectral IV regression misalignment","New loss makes spectral IV learn the right signal","Augmented operator aligns spectral features with the outcome","When spectral features miss, outcome-aware learning recovers","Task-specific spectral features: the cure for misaligned IV"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000138,"raw_usage":{"total_tokens":971,"prompt_tokens":706,"completion_tokens":265,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":189}},"tokens_in":450,"tokens_out":265,"duration_ms":3441,"temperature":1.0,"reasoning_tokens":189,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:18:53.069226+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a finite-rank IV problem where the structural function equals one deep singular function exactly, compute the top singular subspace of Tδ with δ given by the paper's formula, and check whether the singular-value gap is positive and the top left singular vector aligns with the intended direction; then train the one-feature model and compare the L2 error to the bound. If the error does not fall as the signal-to-noise ratio grows with the singular value fixed, the mechanism fails.","supporting_citations":[],"review_version":1}