{"id":"7c9c3433-bcaf-420b-8086-dda5ab84f691","arxiv_id":"2504.13568","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MetaDSE combines MAML meta-training with a workload-adaptive architectural mask to predict CPU IPC and power for unseen workloads from a handful of samples, outperforming transfer-learning baselines on IPC.","lead":"This paper proposes MetaDSE, a meta-learning framework that teaches a CPU performance predictor to adapt to new workloads from only a few simulation samples. On SPEC CPU 2017 it reports up to 44.3% lower prediction error than prior transfer-learning DSE methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WAM's 27% gain lacks a capacity-matched control and Algorithm 2 never updates the mask; the headline 44.3% is IPC-geomean only, not power.","rationale":"The reader's weakest assumption is exactly the load-bearing issue: WAM's benefit could come from fitting extra mask parameters to target data rather than from a transferable architectural property. My read of Algorithm 2 strengthens this: the printed algorithm never updates M, so the paper does not specify whether M is adapted or frozen. If M is adapted, the parameter-count argument applies; if frozen, the mechanism is not workload-adaptive. Either way, the reported 27% needs a capacity-matched control. The 44.3% headline is also overstated as a general prediction-error claim because Table II shows marginal power gains and a power EV regression. These are addressable with additional ablations and clearer reporting, so the conditional verdict stands. I agree with the reader's assessment and recommend no change to the verdict.","tokens_in":11139,"tokens_out":2579,"duration_ms":24414,"concrete_test":"Run an ablation on the five test workloads with identical adaptation steps and learning rates: (a) MetaDSE with M initialized from source attention statistics; (b) same architecture with M randomly initialized; (c) same with all-ones M; (d) MetaDSE-w/o-WAM augmented with a learnable per-head attention scaling vector of the same parameter count. If (b), (c), or (d) reproduces the reported 27% IPC RMSE improvement over MetaDSE-w/o-WAM, the WAM benefit is adaptation capacity, not transferred architectural knowledge. Also re-run Algorithm 2 with M explicitly updated by gradient descent on the support set to determine which version produced the numbers in Table II and Fig. 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a 44.3% prediction-error reduction over TrEnDSE and a 27% gain from WAM. The 44.3% figure is reproduced only by taking the geometric mean of per-workload IPC RMSE in Fig. 5; Table II's averaged IPC RMSE gives 32.6% (0.3270 -> 0.2204), and power RMSE improves only 0.5% while power EV drops from 0.5711 to 0.3189. The claim should therefore be scoped to IPC, with power results reported honestly. More load-bearing is the WAM mechanism. Section IV.C says M is constructed from attention weights of the last self-attention layer aggregated across source workloads, but Algorithm 2 only equips the model with M and sets M.required_grad = True; the inner-loop update rule (line 9) updates theta* only, with no M update shown. If M is frozen during adaptation, the 'adaptive' claim is unsupported; if M is updated on the target support set, the 27% improvement may be extra learnable parameters fitted to target data rather than transferable architectural structure. The paper provides no control distinguishing these, and WAM is the paper's novel transfer contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MetaDSE, a two-stage framework for cross-workload CPU design space exploration: a MAML-based pre-training stage that treats each workload as a task and learns a good transformer-surrogate initialization, and a downstream adaptation stage that uses a novel Workload-adaptive Architectural Mask (WAM) built from attention weights of the last self-attention layer. The paper evaluates MetaDSE on SPEC CPU 2017 against TrEnDSE and other transfer-learning baselines, reporting that MetaDSE reduces prediction error by 44.3% relative to TrEnDSE and that WAM contributes a 27% error reduction. The abstract and introduction present these numbers as the main empirical contributions, alongside an open-source release.","tokens_in":11452,"tokens_out":3840,"duration_ms":36554,"significance":"The core idea of reframing cross-workload CPU DSE as a few-shot meta-learning problem is well motivated and potentially valuable: if the central IPC result holds, the framework reduces the simulation burden for new workloads in a way that is not tied to a workload-similarity assumption. The evaluation on held-out SPEC CPU 2017 workloads is an appropriate external test, and the IPC results in Table II and Fig. 5 provide genuine evidence for the IPC claim. The open-source release is a concrete strength, although I could not verify the artifact during review. The significance is currently undermined by two load-bearing problems: the headline 44.3% reduction is not supported across metrics, and the WAM mechanism is not specified or controlled well enough to support the paper's central novelty claim.","major_comments":[{"comment":"The abstract and contribution statement claim a 44.3% reduction in prediction errors compared with TrEnDSE, but the data in Table II do not support that number as a general claim. For IPC, Table II shows RMSE decreasing from 0.3270 to 0.2204, which is a 32.6% reduction, not 44.3%; the 44.3% figure appears to be the geometric mean of per-workload IPC RMSE in Fig. 5. For power, Table II shows RMSE decreasing from 0.3990 to 0.3969, an improvement of about 0.5%, while explained variance (EV) drops from 0.5711 to 0.3189. This directly contradicts the text in Section VI.A that MetaDSE \"consistently outperforms TrEnDSE and other baselines across all evaluated workloads\" and achieves \"higher EV scores\" on power. The claims should be scoped to IPC geomean RMSE, and the power results should be reported honestly, with an explanation of why WAM adaptation degrades power EV.","section":"Section VI.A, Table II, Fig. 5"},{"comment":"The WAM mechanism is the paper's main novelty, but the algorithm as written does not establish the claimed 27% benefit. Algorithm 2 line 2 sets M to a learnable state, yet the only update rule in the adaptation loop (line 9) updates theta* only and never updates M. If M is frozen during adaptation, the name \"workload-adaptive\" is unsupported; if M is instead updated on the target support set, the update rule is missing. Furthermore, the comparison between MetaDSE and MetaDSE-w/o-WAM does not control for additional parameter capacity: the 27% improvement could come from fitting extra mask parameters to the target support set rather than from transferable architectural structure. The paper needs (a) a precise description of how attention weights from source workloads are aggregated into M, (b) an explicit update rule for M during adaptation or a clear statement that M is fixed, and (c) a capacity-matched control, such as a random fixed mask or a mask learned only from target data with the same number of parameters.","section":"Section IV.C, Algorithm 2"},{"comment":"The evaluation-protocol description is not sufficiently precise to judge whether the results are externally valid. The text says the data were split by \"iteratively and randomly designated seven datasets for training, five for validation, and five for testing,\" but it does not specify which workloads belong to each split, whether the five workloads in Table II are the test workloads, or how the random iterations were generated and aggregated. This matters because Section IV.A states that a meta-validation step is used to select the optimal pre-trained parameters; if the test workloads influence that selection through the random iteration process, the held-out claim is weakened. Please clarify the split protocol, report per-split results, and state whether any model-selection decisions used test-workload information.","section":"Section V, Section VI.A"}],"minor_comments":[{"comment":"Reference [20] is cited twice in the same sentence, and the sentence structure makes the citation ambiguous.","section":"Section II.B"},{"comment":"The legend label \"TrEnDSE-Transfermor\" is misspelled and should be \"TrEnDSE-Transformer\".","section":"Fig. 5"},{"comment":"Line 2 uses \"M.required grad = True\"; this should be \"M.requires_grad_(True)\" or equivalent PyTorch syntax, and the line should indicate whether M is updated in the inner loop.","section":"Algorithm 2"},{"comment":"The MAPE formula uses both y_i and y_i^real without defining their relationship to the notation in Eq. (1); please align the notation.","section":"Section V, Eq. (2)"},{"comment":"The row labeled \"Baseline\" is never defined in the text; it should be explicitly identified (e.g., TrEnDSE) so the comparison is reproducible.","section":"Table III"},{"comment":"The colorbar label reads \"Wasserstein Distance\" while the title reads \"Workload Similarity\"; since a larger Wasserstein distance means lower similarity, the axis semantics should be clarified or inverted to avoid confusion.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible systems/architecture submission with a useful line of work, but the current version overstates its results: the 44.3% headline is not supported across metrics, and the WAM mechanism is both underspecified and uncontrolled. The IPC result may well be salvageable, and the WAM explanation can be fixed with a clearer algorithm and an additional control experiment. If the authors cannot provide a capacity-matched control for WAM or cannot explain the power EV drop, the central novelty claim should be substantially weakened. I also recommend that the editor ask for the code artifact to be checked at revision, since the anonymous link is the basis for the reproducibility claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThis one is worth a look. The core idea—reframing cross-workload CPU DSE as a few-shot meta-learning problem and applying MAML—is sensible and yields a real IPC accuracy gain over TrEnDSE on SPEC 2017. The empirical comparison is new, the code is open-sourced, and the transformer surrogate is a reasonable choice. If the IPC result holds, it removes a practical bottleneck for workload-specific DSE.\n\nWhat's solid: Table II shows IPC RMSE drops from 0.327 to 0.220 (32.6% average), and the Fig. 5 geomean gives the advertised 44.3%. That's a genuine improvement, not a small artifact. The paper also does a decent job situating itself against transfer-learning baselines.\n\nThe soft spots are real but addressable. First, the power claims are not supported. The text says \"consistently outperforms... across all evaluated workloads\" and \"higher EV scores,\" but Table II shows power EV dropping from 0.5711 (TrEnDSE) to 0.3189 (MetaDSE), with RMSE barely improving. The 44.3% should be scoped to IPC, and power needs honest reporting.\n\nSecond, WAM is under-specified and possibly misnamed. Algorithm 2 sets M.required_grad = True but the inner-loop update only touches theta*, so M is never actually updated on the target support set. If M is frozen, \"adaptive\" is wrong; if M is updated, the paper needs to show that update and then control for the extra parameters. The 27% WAM gain comes from an ablation without a capacity-matched control or a random-mask baseline, so the transferable-architecture explanation is not established.\n\nThis is a solid paper with a fixable weakness, not a fatal one. It deserves a serious referee who will push on the WAM details and the power reporting. I'd engage with it.","headline":"Real IPC gains from MAML in cross-workload DSE, but the WAM mechanism is under-specified and the power results are overclaimed.","tokens_in":11956,"tokens_out":2701,"would_cite":false,"duration_ms":22352,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MetaDSE reframes cross-workload CPU design-space exploration as a few-shot meta-learning problem.","keywords":["Design space exploration","Cross-workload prediction","Meta-learning","Few-shot learning","Transformer surrogate","CPU microarchitecture","Workload-adaptive architectural mask"],"falsifier":"Run a controlled ablation in which a mask of the same size and learnability is initialized randomly, rather than from source attention weights, and fine-tuned on the same target support set; if its RMSE matches MetaDSE, WAM is not transferring architectural knowledge. Alternatively, freeze the WAM mask after pre-training so it receives no gradient updates on the target support set, and check whether the 27% improvement remains.","tokens_in":1584,"feed_emoji":"🖥️","tokens_out":1821,"duration_ms":61475,"temperature":0.7,"pith_summary":"This paper argues that CPU design-space exploration across workloads should be treated as a few-shot meta-learning problem, not as transfer learning that assumes workloads are similar. It proposes MetaDSE, in which a transformer-based surrogate is pre-trained with model-agnostic meta-learning (MAML) to produce an initialization that adapts to a new workload with only a handful of simulated samples. For adaptation, it introduces the workload-adaptive architectural mask (WAM), a mask built from attention weights in the transformer's last layer that suppresses irrelevant interactions between architectural parameters. On SPEC CPU 2017, the authors report a 44.3% reduction in prediction error compared with the state-of-the-art TrEnDSE baseline, with WAM accounting for 27 percentage points of that gain. A correct result would substantially lower the simulation budget for evaluating new workloads in CPU design.","feed_headline":"Meta-learning cuts cross-workload CPU prediction error by 44%","feed_subtitle":"A meta-learned transformer plus an architectural mask adapts to a new workload with only five simulator samples.","key_machinery":"The machinery is two-stage. In pre-training, MAML runs an inner loop that fine-tunes the surrogate on each workload's support set and an outer loop that updates the shared initialization using the meta-gradient, thereby learning an initialization that supports rapid adaptation. In adaptation, the workload-adaptive architectural mask (WAM) is built from attention weights of the transformer's last self-attention layer, aggregated over many source workloads to identify frequently recurring parameter interactions; the mask is then attached to the self-attention operator, set learnable, and fine-tuned on the target's support set, filtering out irrelevant interactions while preserving the architectural relationships that transfer.","core_discovery":"The paper's central claim is that the two main weaknesses of cross-workload DSE, overfitting during pre-training and reliance on workload similarity during adaptation, can be addressed by reframing it as few-shot meta-learning. MAML pre-training learns an initialization tailored to fast per-workload adaptation, and WAM adaptation transfers knowledge at the level of architectural parameter interactions rather than workload similarity. As reported, MetaDSE outperforms TrEnDSE across SPEC CPU 2017 workloads: the per-workload geometric mean RMSE drops by 44.3%, and on the five-workload averaged results it achieves instructions-per-cycle (IPC) RMSE of 0.2204 versus 0.3270 and power RMSE of 0.3969 versus 0.3990, with corresponding improvements in MAPE and explained variance.","pith_inferences":["If WAM's 27% gain persists when the mask is held fixed after pre-training and not tuned on target samples, that would confirm it captures transferable architectural structure; if not, the gain is target-specific fitting.","A control experiment that learns a same-size mask from random initialization on the target support set, with the same parameter budget, would isolate how much of WAM's benefit comes from transferable structure versus extra fine-tuning capacity.","The meta-learning formulation could naturally extend to multi-objective design space exploration beyond IPC and power, such as area or timing, though this paper demonstrates only IPC and power on SPEC CPU 2017.","The architectural mask idea suggests a test across design spaces: WAM masks learned on one microarchitecture should transfer to a different microarchitecture if they truly capture architectural parameter interactions."],"forward_implications":["New workloads can be characterized with as few as five simulated configurations, because the meta-learned initialization is designed for quick adaptation.","Workload-similarity metrics such as Wasserstein distance become unnecessary, since the method transfers through architectural structure rather than assuming similar workloads.","A transformer surrogate pre-trained this way beats tree-ensemble transfer baselines on both IPC and power prediction.","Aligning the pre-training support size with the adaptation support size gives the best transfer performance, according to the sensitivity experiments.","The WAM ablation indicates that masking irrelevant parameter interactions is a measurable source of accuracy improvement, not merely a side effect of a larger model."],"supporting_citations":[{"why":"The state-of-the-art cross-workload DSE baseline that MetaDSE is compared against and that the 44.3% reduction is measured relative to.","marker":"[12]"},{"why":"The model-agnostic meta-learning algorithm that supplies the pre-training update rule for learning a fast-adapting initialization.","marker":"[19]"},{"why":"The transformer-based predictor architecture that MetaDSE adopts as its surrogate model.","marker":"[7]"},{"why":"The timing-accurate simulator used to generate the CPU design-space dataset with varied microarchitectural configurations.","marker":"[28]"},{"why":"The power modeling tool used to produce the power target values that MetaDSE predicts alongside IPC.","marker":"[29]"},{"why":"The SPEC CPU 2017 benchmark suite that defines the workloads used for training, validation, and testing.","marker":"[30]"},{"why":"The phase-analysis tool that splits each workload into representative simulation points used in the dataset.","marker":"[31]"}],"fun_headline_variants":["MetaDSE: 44% lower CPU design error via meta-learning","Few-shot meta-learning cuts CPU design error by 44%","MetaDSE: 44% error reduction for cross-workload CPU design","MetaDSE adapts to new CPU workloads with only 5 simulation samples","MetaDSE: few-shot meta-learning for CPU design, 44% error cut"],"cache_read_input_tokens":14080,"weakest_assumption_plain":"The method's load-bearing premise is that a mask built from last-layer attention weights on source workloads captures transferable architectural structure, rather than serving as extra parameters that simply absorb information from the target's few samples; if that premise fails, WAM's reported gain is just additional fine-tuning capacity.","fun_headline_variants_meta":{"raw":{"variants":["MetaDSE: 44% lower CPU design error via meta-learning","Few-shot meta-learning cuts CPU design error by 44%","MetaDSE: 44% error reduction for cross-workload CPU design","MetaDSE adapts to new CPU workloads with only 5 simulation samples","MetaDSE: few-shot meta-learning for CPU design, 44% error cut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001029,"raw_usage":{"total_tokens":4314,"prompt_tokens":905,"completion_tokens":3409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":3310}},"tokens_in":521,"tokens_out":3409,"duration_ms":23479,"temperature":1.0,"reasoning_tokens":3310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:05:19.223296+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled ablation in which a mask of the same size and learnability is initialized randomly, rather than from source attention weights, and fine-tuned on the same target support set; if its RMSE matches MetaDSE, WAM is not transferring architectural knowledge. Alternatively, freeze the WAM mask after pre-training so it receives no gradient updates on the target support set, and check whether the 27% improvement remains.","supporting_citations":[{"cited_title":"A transfer learning framework for high-accurate cross-workload design space exploration of cpu,","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art cross-workload DSE baseline that MetaDSE is compared against and that the 44.3% reduction is measured relative to."},{"cited_title":"Multi-objective optimization in cpu design space exploration: Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"The transformer-based predictor architecture that MetaDSE adopts as its surrogate model."},{"cited_title":"The gem5 simulator,","cited_arxiv_id":null,"evidence_quote":"The timing-accurate simulator used to generate the CPU design-space dataset with varied microarchitectural configurations."},{"cited_title":"Mcpat: An integrated power, area, and timing modeling framework for multicore and manycore architectures,","cited_arxiv_id":null,"evidence_quote":"The power modeling tool used to produce the power target values that MetaDSE predicts alongside IPC."},{"cited_title":"Spec cpu 2017","cited_arxiv_id":null,"evidence_quote":"The SPEC CPU 2017 benchmark suite that defines the workloads used for training, validation, and testing."},{"cited_title":"Simpoint 3.0: Faster and more flexible program phase analysis,","cited_arxiv_id":null,"evidence_quote":"The phase-analysis tool that splits each workload into representative simulation points used in the dataset."}],"review_version":1}