{"id":"14d909fe-ffd8-4b95-a446-044130b12b21","arxiv_id":"2603.23566","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"AscendOptimizer combines kernel rewinding for reusable experience with evolutionary search on hardware feedback to optimize Ascend NPU operators, delivering 1.21x geometric-mean speedup and faster performance on 53.47% of 101 tested operators versus baseline.","lead":"AscendOptimizer is an episodic agent that optimizes AscendC operators for Ascend NPUs by rewinding strong kernel implementations to extract reusable lessons and running profiling-in-the-loop evolutionary search for host-side tiling and data movement. This targets a hardware ecosystem with few public kernels, potentially reducing manual tuning effort for developers working on Ascend NPUs.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Whether rewinding optimizations from strong kernels produces reusable, transferable experience that improves rewriting on the 101 operators (vs. evolutionary search alone) is not isolated in the reported results.","rationale":"Reader's weakest assumption directly matches the load-bearing point. Full-text review confirms the two-part design but does not alter the need for isolation; the concern is internal to the argument rather than external consensus. This moves the verdict from UNVERDICTED to CONDITIONAL pending the ablation.","tokens_in":1779,"tokens_out":327,"duration_ms":30290,"concrete_test":"Run the full pipeline on the 101 operators with the experience-reuse component disabled (no rewinding or experience storage, only the evolutionary search); if geometric-mean speedup falls below 1.1x or loses statistical separation from Best-of-N under identical evaluation budgets, the rewinding mechanism is load-bearing for the central claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline 1.21x geo-mean speedup and outperformance vs. Best-of-N/OpenEvolve rest on the episodic mechanism: controlled removal from strong kernels yields experience that is later reused for rewriting, combined with profiling-in-the-loop evolutionary search for tiling. If the initial strong kernels are obtained only after the evolutionary search has already succeeded, or if the extracted experience does not measurably transfer to dissimilar operators in the benchmark, the gains could be driven primarily by the search component rather than the rewinding+reuse loop. The abstract and method description do not provide an ablation or transfer analysis that separates these contributions.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces AscendOptimizer, an episodic agent for optimizing AscendC operators on Ascend NPUs. It rewinds optimizations from strong kernels in a controlled manner to extract reusable experience for later rewriting, while employing profiling-in-the-loop evolutionary search to discover valid tiling and data-movement configurations on the host side. On a benchmark of 101 real AscendC operators, it reports a 1.21x geometric-mean speedup over the open-source baseline (with 53.47% of operators faster than references) and consistent outperformance versus Best-of-N sampling and OpenEvolve under matched evaluation budgets.","tokens_in":1901,"tokens_out":507,"duration_ms":25698,"significance":"If the central claims hold after addressing the noted gaps, the work offers a practical method for building optimization knowledge in hardware ecosystems with limited public kernels, by combining experience extraction via controlled rewinding with hardware-feedback evolutionary search. This could be relevant for emerging NPUs where traditional tuning resources are scarce.","major_comments":[{"comment":"Abstract and method description: The headline 1.21x geometric-mean speedup and outperformance claims rest on the episodic rewinding+reuse mechanism producing transferable experience that improves rewriting beyond what profiling-in-the-loop evolutionary search achieves alone. No ablation or transfer analysis is provided to isolate this contribution (e.g., comparing full AscendOptimizer against search-only variants on the same 101 operators), leaving open the possibility that gains are driven primarily by the search component rather than the rewinding loop.","section":"Abstract / Method"},{"comment":"Experimental section: The soundness of the performance claims is difficult to evaluate because the manuscript provides no details on experimental controls, statistical testing, operator selection criteria for the 101-operator benchmark, or potential measurement biases (e.g., warm-up, variance across runs). This directly affects the reliability of the reported geometric-mean speedup and fast_p tail ratios.","section":"Experiments"}],"minor_comments":[{"comment":"Clarify the precise definition of 'controlled removal' of optimizations and how 'measurable hurt' is quantified to ensure reproducibility of the experience extraction step.","section":"Method"},{"comment":"Add explicit discussion of invalid or suboptimal trials encountered during evolutionary search and how the budget is allocated across operators.","section":"Method"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify how to better substantiate the contributions of AscendOptimizer. We will revise the manuscript to include a direct ablation isolating the rewinding mechanism and to expand the experimental details for improved reproducibility and reliability assessment.","responses":[{"response":"We acknowledge the value of a direct ablation to isolate the episodic rewinding and experience-reuse component. Our existing comparisons to Best-of-N sampling and OpenEvolve under matched evaluation budgets already show that the full AscendOptimizer outperforms pure search-based approaches on the 101 operators. Nevertheless, to address the referee's point explicitly, we will add an ablation study in the revised manuscript that evaluates a search-only variant (profiling-in-the-loop evolutionary search without the rewinding loop) against the complete system on the identical benchmark, reporting geometric-mean speedup and fast_p ratios. This will clarify the incremental benefit of the rewinding mechanism.","revision_made":"yes","referee_comment":"[Abstract / Method] Abstract and method description: The headline 1.21x geometric-mean speedup and outperformance claims rest on the episodic rewinding+reuse mechanism producing transferable experience that improves rewriting beyond what profiling-in-the-loop evolutionary search achieves alone. No ablation or transfer analysis is provided to isolate this contribution (e.g., comparing full AscendOptimizer against search-only variants on the same 101 operators), leaving open the possibility that gains are driven primarily by the search component rather than the rewinding loop."},{"response":"We agree that additional methodological details are required. In the revised version we will expand the Experiments section with a new subsection that specifies: (i) the criteria used to select the 101 real AscendC operators, (ii) the full measurement protocol including warm-up iterations, number of repeated runs per configuration, and how variance is handled, (iii) the statistical procedures applied (e.g., reporting means with standard deviations and any significance tests), and (iv) controls for hardware and environmental variability. These additions will allow readers to assess the robustness of the 1.21x geometric-mean speedup and the 53.47% fast_p figure.","revision_made":"yes","referee_comment":"[Experiments] Experimental section: The soundness of the performance claims is difficult to evaluate because the manuscript provides no details on experimental controls, statistical testing, operator selection criteria for the 101-operator benchmark, or potential measurement biases (e.g., warm-up, variance across runs). This directly affects the reliability of the reported geometric-mean speedup and fast_p tail ratios."}],"tokens_in":1441,"tokens_out":547,"duration_ms":46794,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this paper shows a practical way to optimize AscendC operators on Ascend NPUs, where public kernels are scarce. It uses an episodic agent that rewinds optimizations from already-strong implementations, keeps the changes that hurt performance as reusable experience, and combines that with profiling-in-the-loop evolutionary search for the host-side tiling and data movement part. On the 101-operator benchmark it reports a 1.21x geometric mean over the open-source baseline and beats Best-of-N and OpenEvolve at equal evaluation budgets, with 53% of operators improving.","headline":"AscendOptimizer gets a 1.21x geo-mean speedup on 101 AscendC operators by rewinding strong kernels for experience and running hardware-feedback evolutionary search, but the results do not isolate how much the rewinding step adds over search alone.","tokens_in":2424,"tokens_out":213,"would_cite":false,"duration_ms":27684,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"AscendOptimizer episodic rewind-and-reuse agent for NPU kernels shares no machinery with RS forcing or J-cost","alignment":"orthogonal","rationale":"Paper centers on LLM-driven evolutionary tiling search plus optimization-rewind to synthesize bad-to-good trajectories distilled into a RAG experience bank (Stage I/II alternating loop, Algorithm 1). RS derives c, ℏ, G, 3D spacetime and 8-tick periodicity parameter-free from a single distinction via J(x)=½(x+x⁻¹)−1 and φ-ladder (reality_from_one_distinction, AbsoluteFloorClosure, Cost.FunctionalEquation). No ratio-symmetric cost, golden-ratio identities, or recognition-cost forcing appears; domain is practical systems optimization with zero overlap.","tokens_in":54464,"confidence":"high","tokens_out":171,"duration_ms":14223,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"AscendOptimizer is an episodic agent that learns Ascend NPU operator optimizations by rewinding proven kernels to extract reusable experience and by running profiling-driven evolutionary search for tiling and data movement.","keywords":["Ascend NPU","AscendC operators","episodic agent","operator optimization","tiling configurations","evolutionary search","profiling in the loop","kernel rewriting"],"falsifier":"On a new held-out set of AscendC operators, AscendOptimizer fails to produce higher geometric-mean speedup than Best-of-N sampling or OpenEvolve when given identical per-operator evaluation budgets.","tokens_in":2678,"feed_emoji":"","tokens_out":681,"duration_ms":37610,"temperature":0.7,"pith_summary":"The paper addresses the scarcity of public optimization examples for AscendC operators on Ascend NPUs, where performance hinges on both a host-side tiling program and the kernel itself. It builds missing knowledge directly from hardware execution rather than external datasets. For kernels, the agent removes optimizations from strong implementations in a controlled manner and retains only the removals that measurably degrade speed as reusable rewriting experience. For the host side, it couples evolutionary search with on-device profiling to discover valid, high-performance tiling configurations. On 101 real operators this yields a 1.21 times geometric-mean speedup over the open-source baseline while outperforming Best-of-N sampling and OpenEvolve under equal evaluation budgets.","feed_headline":"Episodic agent speeds Ascend NPU operators 1.21x","feed_subtitle":"By extracting experience from rewound kernels and profiling tiling search, it beats sampling baselines on 101 real operators.","key_machinery":"The episodic agent combining controlled optimization removal for kernel rewriting experience with profiling-in-the-loop evolutionary search for tiling configurations.","core_discovery":"AscendOptimizer is an episodic agent that improves AscendC operators by rewinding optimizations from strong implementations to create reusable experience and by performing profiling-in-the-loop evolutionary search for host-side tiling and data-movement configurations, leading to a 1.21x geometric mean speedup over the open-source baseline on 101 operators with 53.47 percent outperforming their references.","pith_inferences":["The rewind-and-reuse pattern could transfer to operator optimization on other hardware platforms that also lack public high-performance examples.","Accumulated experience across many operators might eventually lower the number of evaluations needed for each new operator.","The approach could be extended to discover operator fusions or alternative pipelining strategies not present in the original references."],"forward_implications":["Kernel structure and host-side scheduling can be improved jointly without large external kernel databases.","Performance gains appear consistently across different evaluation budgets and outperform simple sampling baselines.","More than half the optimized operators exceed the speed of their original reference implementations.","The same workflow can be applied to additional AscendC operators beyond the 101-operator benchmark."],"fun_headline_variants":["Episodic rewinds accelerate Ascend NPU operators to 1.21x","Kernel rewinds and tiling search speed Ascend NPU to 1.21x","AscendOptimizer uses rewinds for 1.21x on 101 AscendC operators","Rewinding and evolutionary search speed Ascend NPU operators 1.21x"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Removing optimizations from strong kernels in a controlled way yields reusable, generalizable experience that helps rewrite new operators, and profiling-guided evolutionary search finds valid high-performance tiling configurations within modest evaluation budgets.","fun_headline_variants_meta":{"raw":{"variants":["Episodic rewinds accelerate Ascend NPU operators to 1.21x","Kernel rewinds and tiling search speed Ascend NPU to 1.21x","AscendOptimizer uses rewinds for 1.21x on 101 AscendC operators","Rewinding and evolutionary search speed Ascend NPU operators 1.21x"]},"model":"grok-4.3","cost_usd":0.01845,"raw_usage":{"total_tokens":7828,"prompt_tokens":681,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":184499500,"prompt_tokens_details":{"text_tokens":681,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":7064,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":681,"tokens_out":83,"duration_ms":74485,"temperature":1.0,"reasoning_tokens":7064,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-21T10:11:55.049098+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a new held-out set of AscendC operators, AscendOptimizer fails to produce higher geometric-mean speedup than Best-of-N sampling or OpenEvolve when given identical per-operator evaluation budgets.","supporting_citations":[],"review_version":1}