{"id":"c9af69cd-4ef9-472d-af2f-5a5ffb0a1687","arxiv_id":"2608.05571","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A training-free closed-form alignment of retrieved windows outperforms a trained fusion adapter on frozen time-series foundation models.","lead":"Align-RAG shows that a frozen time-series foundation model can use retrieved examples as in-context demonstrations without any learned adapter, if the examples are first rescaled and phase-shifted to match the query. On a standard seven-dataset benchmark it beats a state-of-the-art trained retrieval-fusion model on every dataset in MSE, with an average 3.75% improvement.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 7/7 headline is partly a product of hyperparameters selected on the test benchmark (α=0.60, Wiener-vs-clipped); a validation-based choice could reduce the win count, and the pure aligned-ICL variant already wins only 6/7.","rationale":"The reader's weakest assumption correctly identifies test-set hyperparameter selection, but also assigns weight to the future-shuffle confound. I think the hyperparameter concern is the more load-bearing of the two because it targets the paper's headline empirical claim (7/7, avg −3.75%), while the shuffle confound only weakens the supporting mechanistic interpretation. The paper has real independent support: the 2×2 ablation shows alignment matters (no-align regresses 20+ pp), the pure aligned-ICL variant still beats TS-RAG on 6/7, bootstrap CIs are used for most small datasets, and all code and configurations are released. However, Table 6's selection of α=0.60 from test-set sweeps, combined with Table 7's pure-variant 6/7 result, makes the all-seven claim fragile under an honest validation protocol. The paper should add a validation-based selection check or soften the headline. This keeps the verdict CONDITIONAL, matching the reader's conditional verdict without requiring a change.","tokens_in":17611,"tokens_out":8431,"duration_ms":75229,"concrete_test":"Hold out the first 20% of each test set (by time order) as validation, select α∈{0.50,...,0.70} and the amplitude regularizer (Wiener vs clipped) by average MSE on those validation windows, and then recompute Table 1 on the remaining test windows. If the validation-selected configuration still gives 7/7 MSE wins and an average gain near −3.75%, the concern is resolved; otherwise the headline should be downgraded to the pure aligned-ICL result or reported as a distribution over validation splits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 6 in Appendix C reports an α sweep on the full test sets; α=0.60 is released because it yields 7/7 MSE wins and 6/7 MAE wins, while α=0.62 drops to 6/7 MSE wins. Appendix F similarly selects Wiener shrinkage over clipped affine after inspecting cross-backbone test results on Weather and Exchange (Table 8). The phrase 'no per-backbone tuning' means no learned parameters, but the configuration is tuned at the benchmark level on the test sets. The headline −3.75% average is the selected α's average, not a parameter-free prediction. The selection matters because the all-seven property is not intrinsic to alignment: Table 7 shows the pure aligned-ICL variant (α=β=0) wins only 6/7 datasets and loses on Electricity (+4.1% vs TS-RAG). Thus the strongest claim as stated—'outperforms on all seven datasets'—is conditional on test-selected post-processing weights. A validation-based choice of α and the regularizer would settle whether 7/7 survives.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Align-RAG, a training-free retrieval-augmented forecasting method that aligns retrieved past–future windows to the query using a closed-form Wiener-style amplitude rescaling and an integer-lag phase shift before packing them into a frozen TSFM's context. The method also includes MMR diversification, a tiered token-budget layout, a distance-weighted future blend, and a second consensus forward pass. Against the trained TS-RAG adapter on a frozen Chronos-Bolt, the authors report lower MSE on all seven benchmark datasets (average −3.75%) and lower MAE on six of seven, with paired moving-block bootstrap CIs on five of six small datasets. The same configuration improves zero-shot MSE on four additional frozen TSFMs. A mechanism analysis compares the backbone's prediction shift to a closed-form ridge predictor and reports a future-shuffle control. The paper argues that learned fusion is unnecessary and that closed-form alignment should be the default baseline.","tokens_in":17846,"tokens_out":6370,"duration_ms":54733,"significance":"If the central claim survives scrutiny, the result is significant: it would show that gains previously attributed to trained fusion modules in retrieval-augmented forecasting can be recovered by a parameter-free alignment step, and it would provide a strong, reproducible baseline for the field. The manuscript has notable strengths: the main comparison uses TS-RAG's released retrieval index, the paired moving-block bootstrap is appropriate, code and seeds are released, the comparison against RAFT at multiple horizons is a useful addition, and the cross-backbone results cover several architectures. The limitations are also acknowledged honestly. However, the headline 'all seven datasets' claim is currently supported only by a configuration whose key hyperparameters (α, the Wiener-vs-clipped regularizer, and other post-processing choices) are selected after inspecting the test benchmark, and the pure aligned-ICL variant wins on only 6/7 datasets. These issues do not invalidate the method but they do mean the paper's strongest claim is not yet established as a parameter-free prediction.","major_comments":[{"comment":"The α sweep is computed on the full test sets and is used to pick the released value α=0.60, which yields the 7/7 MSE win count; the adjacent value α=0.62 yields only 6/7 MSE wins. This makes the all-seven headline a post-selection statement, not a parameter-free prediction. In addition, the text in Appendix C states that the MSE-win count is 7/7 for every α∈[0.50,0.62], which is directly contradicted by Table 6 showing α=0.62 with 6/7 MSE wins. Please resolve this inconsistency and re-run the comparison with α chosen on a validation split, reporting both the validation-selected setting and the test results under that setting.","section":"Appendix C, Table 6; §3.3"},{"comment":"The choice of Wiener shrinkage over clipped affine regularization is made after inspecting cross-backbone test results on Weather and Exchange, and the Wiener row reproduces the deployed values from Table 3. This is benchmark-level selection on the test sets for a component that is presented as configuration-free. The claim 'no per-backbone tuning' is technically true only in the sense that the same configuration is reused across backbones; it does not mean the configuration was not tuned on the test benchmark. Please report the sensitivity of the headline win counts to a validation-based choice of this regularizer, or clearly state that Wiener shrinkage was selected on the test benchmark.","section":"Appendix F, Table 8; §4.4"},{"comment":"The future-shuffle control is confounded by the future-blend term. The paper's full prediction uses by_u = (1−β)f_θ(C) + β Σ_i w_i ˜f_i, and the second pass by_c contains the same blend. If the shuffled futures are used in this blend, permuting {f_i} changes the weighted sum Σ_i w_i f_i when the weights differ, so the observed MSE reversal under the shuffle may reflect a changed blend prior rather than the frozen backbone's in-context regression on corrupted pairs. Please run the shuffle with the blend term held fixed (e.g., using the pure aligned-ICL variant with β=0, or explicitly reusing the unshuffled blend weights and futures) to isolate the context effect, and state whether by_c also changes under the shuffle.","section":"§5, Eq. (3), Figure 2"},{"comment":"The abstract's claim of outperforming TS-RAG 'on all seven datasets' is carried by the full configuration with β=0.15 and the α=0.60 consensus pass. Table 7 shows that the pure aligned-ICL variant (β=0, α=0) wins on only 6/7 datasets and loses on Electricity (+4.1% MSE). Since α and β are selected on the test benchmark, the all-seven property is not intrinsic to alignment alone. Please qualify the headline claim accordingly, and report win counts for the pure aligned-ICL variant as the primary evidence for the 'alignment is sufficient' thesis.","section":"Table 7; Abstract; §3.3"}],"minor_comments":[{"comment":"The same integer lag τ* is applied to the retrieved future, but the boundary convention is not specified: with |τ| up to S/4=128 and future length H=64, shifting the future by a nonzero lag changes its length. Please state how the shifted future is truncated or padded to remain length H.","section":"§3.2, Eq. (2)"},{"comment":"The phrase 'with no learned parameters' is accurate but could be misread as 'no hyperparameters.' The method has several hand-set hyperparameters (K, λ, β, α, M, layout, τ search range, distance temperature), and Appendix F shows that at least one of them is chosen on the test benchmark. Please use 'no trained parameters' or 'training-free' consistently and list which hyperparameters are set a priori versus selected on test data.","section":"Abstract; §3.1; Appendix A"},{"comment":"The abstract's 'all seven datasets' phrasing should be qualified by the paper's own significance statements: Exchange's bootstrap CI includes zero and Electricity is not tested with a CI. Suggest wording such as 'lower MSE point estimate on all seven datasets, with significant gains on five of six small datasets' to avoid overclaiming.","section":"Table 1; §4.2"},{"comment":"The text says 'four additional frozen TSFMs' while Table 3 includes five backbones: Bolt, Chronos-2, TimesFM, Moirai, and Toto. The intended meaning is four backbones beyond Chronos-Bolt, but the wording should be clarified to avoid confusion.","section":"§4.4; Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and the core idea is appealing, but the strongest claim (7/7 unconditional wins with no tuning) is currently weakened by test-set selection of α and the regularizer. The good news is that the issue is fixable within the manuscript's scope: a validation-based hyperparameter choice, a corrected α-sweep table, and a deconfounded future-shuffle control would address all major concerns. The code release and the use of TS-RAG's own retrieval index are valuable assets for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The first thing to know: the paper's core claim is that a completely training-free, closed-form alignment of retrieved windows (amplitude + phase) beats the trained TS-RAG adapter on a frozen Chronos-Bolt on 7/7 datasets. That claim is only true for the full configuration, which includes a second-pass average and a future blend whose weights are selected on the test sets. The pure aligned-ICL variant, with those terms removed, wins 6/7 and loses on Electricity. The paper discloses this, but the abstract's 7/7 headline reads as unconditional.\n\nWhat is actually new: applying amplitude rescaling and integer-lag phase alignment to retrieved past–future windows before they enter a frozen TSFM, and showing this makes a trained fusion module unnecessary. The 2x2 ablation cleanly separates alignment from retrieval ranking and shows alignment carries the gain. The bootstrapped paired comparisons are careful. The cross-backbone table and multi-horizon consistency support the empirical finding. Code, data, and seeds are released. This is a useful, reproducible result.\n\nThe soft spots are real but not fatal. Appendix C sweeps α on the full test sets and picks 0.60, the value giving 7/7 MSE and 6/7 MAE. Table 6 shows α=0.62 already drops to 6/7 MSE, which contradicts the text's claim that the win count is 7/7 across [0.50,0.62]. Appendix F picks Wiener shrinkage after comparing on test-set Weather and Exchange; on Exchange, Wiener is worse than clipping for TimesFM. So the configuration is benchmark-level tuning on test data. A validation-based choice is needed to know whether the 7/7 survives.\n\nThe mechanism section has a specific confound. The prediction blends the model output with a weighted average of the retrieved futures (β term). When they shuffle the futures for the control, they change that blend term directly. The observed MSE reversal and preserved ridge-signature could come from the corrupted blend, not from the frozen backbone's in-context behavior. A β=0 future-shuffle control is needed to make the causal claim. The ridge-tracking idea is interesting, but the current evidence doesn't isolate it.\n\nWho is this for: anyone working on retrieval-augmented forecasting or TSFM adaptation. It provides a strong free baseline that should be tried before training adapters. The paper deserves a serious referee—the central finding is important and likely correct, but the test-set selection and the confound need to be addressed in revision. I'd accept it for review and would press hard on those two points.","headline":"The 7/7 headline is partly an artifact of test-set hyperparameter selection, but the core finding—training-free alignment beats trained fusion—is real and worth a close look.","tokens_in":18420,"tokens_out":5211,"would_cite":true,"duration_ms":39574,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"With no learned parameters, aligned retrieved examples beat a trained fusion adapter on 7/7 benchmarks and 4 more backbones.","keywords":["retrieval-augmented forecasting","time series foundation models","in-context learning","training-free alignment","closed-form ridge regression","zero-shot forecasting","frozen backbone","amplitude and phase alignment"],"falsifier":"Re-run the headline comparison with $\\alpha$ and $\\beta$ chosen on a validation split of each dataset instead of on the test sweep; if the 7/7 MSE wins shrink or reverse, the average −3.75% gain was partly test-set selection.","tokens_in":17399,"feed_emoji":"📈","tokens_out":9360,"duration_ms":72797,"temperature":0.7,"pith_summary":"Align-RAG claims that a frozen time-series foundation model can use retrieved examples effectively without any learned fusion module, provided the retrieved past–future windows are first aligned to the query in amplitude and phase by closed-form transforms. On the standard seven-dataset benchmark with a frozen Chronos-Bolt, this training-free method beats the trained TS-RAG retrieval adapter on all seven datasets in MSE (average −3.75%) and on six of seven in MAE. The same fixed configuration improves zero-shot MSE on four additional frozen backbones by 2.5% to 13.7% without per-backbone tuning. A retriever-by-alignment ablation attributes the gain to alignment rather than retrieval ranking, and a behavioral probe finds the backbone's prediction shift under aligned demonstrations tracks a closed-form ridge predictor. If right, this means the learned fusion modules in current retrieval-augmented forecasting are largely compensating for missing alignment, and closed-form alignment should be the default baseline.","feed_headline":"Training-free alignment beats trained adapter on 7/7 datasets","feed_subtitle":"Closed-form amplitude and phase alignment recovers the gains of a trained retrieval mixer, with no learned parameters.","key_machinery":"The load-bearing machinery is Align-RAG's closed-form demonstration alignment, applied before the backbone tokenizes the context. For each retrieved past–future pair $(p_i, f_i)$, a regularized affine map with slope $a_i = \\frac{\\sigma_q \\sigma_{p_i}}{\\sigma_{p_i}^2 + (\\sigma_q/M)^2}$ and intercept $b_i = \\mu_q - a_i \\mu_{p_i}$ rescales both past and future so the past matches the query's first two moments; then the pair is shifted by the integer lag $\\tau_i^*$ that maximizes its sample cross-correlation with the query past, with $|\\tau| \\le S/4$, applied identically to the future so the pair stays consistent. Diversification (Maximal Marginal Relevance), a tiered token-budget layout, a distance-weighted future blend, and a no-demonstration second-pass average complete the pipeline, but the pure aligned in-context variant (without blend or second pass) already beats the trained fusion baseline on 6/7 datasets. The mechanism probe uses the closed-form ridge predictor $\\hat{y}_{\\mathrm{ridge}} = q^\\top(P^\\top P + \\lambda I)^{-1} P^\\top F$ on the same aligned pairs as a behavioral reference for the frozen backbone's prediction shift.","core_discovery":"On the paper's own terms, the central discovery is that the performance gains attributed to learned fusion in retrieval-augmented time-series forecasting are recoverable without any training: a frozen Chronos-Bolt, given retrieved (past, future) windows that have been rescaled by a per-pair regularized affine map and shifted by an integer-sample lag chosen by cross-correlation, outperforms TS-RAG's trained Adaptive Retrieval Mixer on 7/7 datasets in MSE (average −3.75%) and 6/7 in MAE. The same recipe, with identical hyperparameters, improves zero-shot MSE on Chronos-2, TimesFM-2.0, Moirai, and Toto by 2.5% to 13.7% per backbone, with no per-backbone tuning. The paper further argues the direction of causality: switching the trained retriever for random retrieval costs under one percentage point of average MSE when alignment is kept, whereas removing alignment costs over 20 percentage points, locating the gain in demonstration alignment rather than retrieval ranking. A behavioral probe shows the model's prediction shift under aligned demonstrations tracks a closed-form ridge predictor on the same pairs, and a future-shuffle control that breaks pair correspondence reverses the gain, consistent with the model regressing on aligned (past, future) pairs rather than averaging their futures.","pith_inferences":["If alignment is the main driver, retrieval-augmented forecasting may be better understood as a normalization problem than a fusion problem; future systems might replace trained adapters with learned or adaptive alignment and still improve.","The ridge-tracking result suggests that even frozen TSFMs behave like implicit linear regressors on aligned in-context pairs; a direct test would be to fit the ridge predictor on random aligned pairs and check whether the backbone's forecast matches on held-out queries, not just in the shift cosine.","Because the reported $\\alpha$ and $\\beta$ sweeps are computed on the test sets, a validation-split selection of these weights would sharpen the claim; if the 7/7 wins persist under that protocol, the training-free advantage is more clearly established.","The future-shuffle control changes the future-blend term along with the pair correspondence; re-running the control with $\\beta=0$ would isolate the frozen backbone's behavior from the blending prior."],"forward_implications":["Learned fusion adapters are not necessary for retrieval-augmented forecasting on frozen TSFMs; a closed-form alignment step should become the standard baseline before training any fusion module.","Gains previously attributed to trained mixers can be recovered without training, so deployment cost drops: no training data, no per-backbone tuning, and inference stays a single forward pass plus lightweight transforms.","Alignment, not retrieval ranking, is the dominant factor: random retrieval with alignment still beats the trained mixer on 4/7 datasets, while removing alignment costs more than 20 points of MSE.","Frozen time-series backbones can use retrieved context dynamically, and their behavior under aligned demonstrations is consistent with an implicit ridge regression on the (past, future) pairs, not with averaging retrieved futures."],"supporting_citations":[{"why":"Supplies the trained fusion baseline (TS-RAG), the seven-dataset benchmark protocol, and the released retrieval index used in the head-to-head comparison.","marker":"[Ning et al., 2025]"},{"why":"Defines Chronos-Bolt, the frozen backbone for the main comparison.","marker":"[Ansari et al., 2024]"},{"why":"Prior work that re-pretrains a TSFM to consume in-context retrieval examples; the approach Align-RAG argues against for the frozen case.","marker":"[Faw et al., 2025]"},{"why":"Shows transformers can implement closed-form ridge regression in-context, the behavioral reference for the paper's mechanism probe.","marker":"[von Oswald et al., 2023]"},{"why":"Establishes that transformers trained on linear functions match least-squares in-context, underlying the implicit-regression interpretation.","marker":"[Garg et al., 2022]"},{"why":"Motivates the regularized shrinkage in the amplitude-alignment slope.","marker":"[Wiener, 1949]"},{"why":"Provides the maximum-likelihood lag estimator used for integer phase alignment.","marker":"[Knapp and Carter, 1976]"},{"why":"Defines Maximal Marginal Relevance used for demonstration diversification.","marker":"[Carbonell and Goldstein, 1998]"}],"fun_headline_variants":["No training, closed-form alignment beats learned fusion","Zero-learned-param alignment outperforms trained adapter on 7/7","Training-free alignment: gains of fusion without the training","Align-RAG: alignment suffices, zero training required","Closed-form alignment beats trained fusion with zero training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline average assumes the fixed hyperparameters (notably the blend weight $\\beta=0.15$ and the second-pass weight $\\alpha=0.60$) were not effectively chosen on the test benchmark, since the reported sensitivity sweeps are computed on the test sets.","fun_headline_variants_meta":{"raw":{"variants":["No training, closed-form alignment beats learned fusion","Zero-learned-param alignment outperforms trained adapter on 7/7","Training-free alignment: gains of fusion without the training","Align-RAG: alignment suffices, zero training required","Closed-form alignment beats trained fusion with zero training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00083,"raw_usage":{"total_tokens":3719,"prompt_tokens":1136,"completion_tokens":2583,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":752,"completion_tokens_details":{"reasoning_tokens":2503}},"tokens_in":752,"tokens_out":2583,"duration_ms":22904,"temperature":1.0,"reasoning_tokens":2503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:16:18.413588+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the headline comparison with $\\alpha$ and $\\beta$ chosen on a validation split of each dataset instead of on the test sweep; if the 7/7 MSE wins shrink or reverse, the average −3.75% gain was partly test-set selection.","supporting_citations":[],"review_version":1}