{"id":"511c66a5-3b4b-418f-909a-583dcd3828bb","arxiv_id":"2508.00806","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"Adacc adaptively chooses between activation recomputation, retention, and compression at tensor level to improve LLM training throughput by 1.01x-1.37x without accuracy loss.","lead":"Adacc is a memory-saving framework for training large AI models that decides, for each piece of data, whether to store it, compress it, or recalculate it later. It claims to speed up training by up to 37% while keeping accuracy about the same.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The provided body is corrupted/unreadable beyond the abstract, so the claimed 1.01–1.37x throughput gain is currently unverifiable; that is a verification gap, not evidence of failure.","rationale":"The reader marked the paper UNVERDICTED because only the abstract was intact; I agree. The weakest point is not one equation or parameter but the absence of any checkable evidence for the headline result. Since the manuscript body is unreadable, any specific technical objection (e.g., to the MILP cost model) would be speculative: I cannot confirm the form of the cost model, the objective, or the experimental protocol. The exception is the concrete anomaly that the rendered text contains a different arXiv ID and repeated equations, which independently supports the conclusion that the current text cannot bear weight. Under the page-facing rule, this is flagged as a limitation of the submitted material itself. I am not accusing the authors of anything; an encoding artifact could be responsible. The correct status is unverified: if a clean version shows the experiments and cost-model validation, the paper could be reviewable; if not, the central claim remains unsupported. A targeted check on MILP overhead is the most decisive because the 1.01–1.37x range is small enough that scheduling overhead could consume the entire benefit; this is also close to the reader's concern, so I partially agree.","tokens_in":9032,"tokens_out":4542,"duration_ms":56203,"concrete_test":"Obtain the clean source or PDF of arXiv:2508.00806 and locate the experiments section. Then run the single decisive check: recompute the reported speedup for the largest model with MILP solver time included in end-to-end throughput. If the speedup disappears or falls below 1.0 when solver time is counted, or if the comparison baselines use different memory budgets or training-step counts, the central claim fails. If the clean paper contains full tables with wall-clock times, MILP overhead, and accuracy-per-token comparisons, this concern is resolved and review can proceed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Adacc's central claim is an empirical one: tensor-level choices among recomputation, retention, and compression yield 1.01–1.37x throughput over state-of-the-art frameworks with baseline-comparable accuracy. For this to be supportable, the MILP scheduler's cost model must be accurate enough that its plan is genuinely near-optimal, the layer-specific compression must preserve accuracy at scale, and the adaptive policy-evolution mechanism must not consume the savings it creates. None of these can be checked in the submitted text: the full text is an unintelligible stream, and a mid-document header 'arXiv:2508.00812v2 [math.OC] 25 Aug 2025' indicates the rendered body is not the paper's actual content. There are no readable tables, model sizes, baseline definitions, wall-clock timings, MILP solve times, or ablations; no formal verification or reproducibility artifact is present. The abstract alone is too thin to establish the central claim, and the concern is therefore unfalsifiability from the provided material, not an identified internal error.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adacc, an adaptive memory optimization framework for training large language models that unifies activation recomputation and data compression. Based on the abstract, Adacc makes tensor-level decisions among recomputation, retention, and compression, using layer-specific compression algorithms, a MILP-based scheduling policy, and an adaptive policy evolution mechanism. The claimed results are a training throughput improvement of 1.01x to 1.37x compared to state-of-the-art frameworks while maintaining baseline-comparable accuracy. However, the submitted full text is an unreadable corrupted stream that contains no complete sentence, equation, table, or experimental result; thus the summary is based solely on the abstract and the methodology is not assessable from the provided material.","tokens_in":9240,"tokens_out":3478,"duration_ms":38291,"significance":"If the claims are substantiated, Adacc could be a practically valuable contribution to memory-efficient LLM training, as it would provide a unified, adaptive alternative to static recomputation or naive compression. The abstract suggests a principled combination of techniques, and the reported throughput gains, if real and reproducible, would be of interest to the machine learning systems community. However, the significance cannot currently be judged: the manuscript contains no readable technical content, no experimental derivation, no baseline definitions, and no evidence such as results tables or error bars. The paper also appears to make a 'first' claim, which requires a careful literature comparison that is absent from the readable portion. The lack of a verifiable body makes any assessment of novelty, correctness, or empirical support impossible at this stage.","major_comments":[{"comment":"The entire body of the manuscript is an unintelligible stream of corrupted characters. There is a mid-document header 'arXiv:2508.00812v2 [math.OC] 25 Aug 2025', which is not this paper's identifier and indicates that the rendered content is not the actual paper. Because no equations, algorithm pseudocode, tables, or figures are readable, the central empirical claim (throughput improvement of 1.01x–1.37x with maintained accuracy) is completely unverifiable. This is a load-bearing issue: the paper's contribution is empirical, and the submission provides no evidence to evaluate.","section":"Full Text"},{"comment":"The reported throughput range of 1.01x to 1.37x is presented without any experimental context. The abstract does not specify the LLM architectures, model sizes, baseline frameworks, hardware, memory constraints, or the number of training tokens. The lower bound of 1.01x is close to noise, and without standard deviations or repeated runs the practical significance of the improvement cannot be assessed. The abstract alone is insufficient to establish the claimed advantage.","section":"Abstract"},{"comment":"The claim that the MILP-based policy 'globally optimizes memory strategies across layers' depends on the accuracy of cost models for memory footprint, compression ratio, and recomputation time for each tensor. The abstract provides no validation of these cost models or any measurement of MILP solver runtime. If the cost models are inaccurate, the optimal plan will not yield the claimed throughput gains, and the solver overhead itself could negate the benefits. This is a central dependency that the unreadable full text cannot support.","section":"Abstract (MILP scheduling policy)"}],"minor_comments":[{"comment":"The term 'retention' is used as one of the tensor-level options, but its precise meaning is not defined in any readable part of the manuscript; the authors should provide a formal definition in the revised text.","section":"Abstract"},{"comment":"The description 'the first adaptive memory optimization framework that unifies activation recomputation and data compression' is a strong claim that requires a thorough related-work comparison; the citation context cannot be evaluated from the submitted material.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The submitted text appears to be corrupted or mis-rendered beyond the abstract. The document contains a header from a different arXiv paper and no readable technical content. I recommend requiring the authors to submit a clean, complete manuscript before any further review. The abstract alone is far too thin to judge the paper's claims, and no amount of refereeing can address the current lack of evidence. This is a submission integrity issue rather than a scientific flaw, so I am not recommending rejection, but the manuscript cannot be reviewed in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou asked for a take on arXiv:2508.00806. The abstract describes a sensible framework, but the full text is a corrupted encoding, not the paper. A header for a different arXiv paper (math.OC) appears mid-document. So I can only judge the abstract, and everything below is provisional on the real manuscript being readable.\n\nWhat's new: Adacc unifies activation recomputation and data compression at tensor level, with a MILP-based scheduler and an adaptive policy-evolution mechanism. That combination is not something I've seen done exactly this way. Prior work mostly uses one technique or static rules. The layer-specific outlier-aware compression is a reasonable touch, since LLM activations have outliers that naive compression mishandles. If the claims hold, this is a moderate but real contribution.\n\nWhat's good: The design is coherent and the problem is real—memory is often the bottleneck in LLM training. The claimed 1.01–1.37x throughput improvement is modest enough to be plausible; it's not a leap, which makes it more believable than a flashy 2x claim.\n\nSoft spots: First, the obvious one—the full text is unreadable, so the experimental section, tables, and ablations are missing. I cannot verify the central claim, and neither can anyone else from this submission. That's a verification gap, not proof of failure, but it's a hard gap. Second, even from the abstract, the MILP scheduling policy depends on accurate cost models for memory, compression ratio, and recomputation time. The abstract gives no evidence these models are accurate, nor any measurement of solver overhead. If the cost models are off, the \"globally optimal\" plan may not deliver the gains. Third, there are no details on model sizes, baselines, hardware, or error bars. The 1.01x lower bound is suspiciously close to \"no improvement\"—I'd want per-configuration results to see if the gain is consistent or just on specific setups.\n\nNet: The idea is plausible and worth a serious look, but only if the authors provide a clean, complete manuscript. The current version is not reviewable. My recommendation: ask the authors for a corrected PDF. If the body matches the abstract and includes the missing experimental detail, send it to referees. If not, desk reject. I'd want to see an ablation on the MILP cost model accuracy and the adaptive policy overhead before believing the throughput numbers.\n\nThat's my read. Happy to discuss.","headline":"The abstract is plausible and worth reviewing, but the submitted body is corrupted and unreadable, so the 1.01-1.37x claim is currently unverifiable.","tokens_in":9727,"tokens_out":3209,"would_cite":false,"duration_ms":36565,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adacc claims LLM training memory can be managed per tensor: retain, compress, or recompute each activation, cutting memory cost while preserving accuracy and delivering 1.01x-1.37x throughput.","keywords":["LLM training","activation recomputation","activation compression","memory optimization","MILP scheduling","outlier-aware compression","adaptive policy","training throughput"],"falsifier":"Take a transformer whose activations have no prominent outlier channels and run Adacc against a pure selective-recomputation baseline under the same memory budget; if throughput is not better and accuracy is worse, the compression branch is not pulling its weight. More directly, perturb the per-tensor memory and recomputation-time estimates fed into the MILP by 20% and compare predicted versus measured iteration time—if the plan's advantage tracks the perturbed estimates, the system is optimizing its own cost model rather than actual runtime.","tokens_in":8868,"feed_emoji":"🧠","tokens_out":5929,"duration_ms":70679,"temperature":0.7,"pith_summary":"Training large language models on GPUs is often blocked by memory, and the two standard workarounds—recomputing activations in the backward pass or compressing them once stored—each exact a price: extra compute or lost accuracy. This paper claims that price can be avoided by treating the choice as a per-tensor optimization problem rather than a fixed policy. Adacc decides tensor-by-tensor whether to keep, compress, or recompute each activation, selects layer-specific compression tuned to activation outliers, and re-plans periodically as training statistics drift. The reported result is 1.01x-1.37x higher training throughput than state-of-the-art frameworks with accuracy comparable to the uncompressed baseline.","feed_headline":"Adaptive recompute-or-compress choices lift LLM training speed 1.37x","feed_subtitle":"Tensor-by-tensor keep, compress, or recompute choices preserve accuracy while cutting memory pressure.","key_machinery":"The load-bearing object is the per-tensor action decision: for each activation tensor, exactly one of three actions is chosen—retain, compress, or recompute. A mixed-integer linear program (MILP) takes the memory footprint, compression ratio, recomputation cost, and outlier structure of every tensor and selects the combination that minimizes expected training time under a global memory budget. Complementing that scheduler are layer-specific compression algorithms that protect outlier channels, which dominate reconstruction error in LLM activations, and a policy-evolution loop that refreshes the plan when the data distribution shifts. Together these pieces convert memory savings into a continuous optimization over training steps.","core_discovery":"Adacc's central claim is that activation recomputation and activation compression are complementary levers, not competing alternatives, and that their best mix is different for every tensor and every phase of training. On the paper's account, a global optimizer should be able to spend a limited memory budget by retaining some tensors untouched, compressing others with an outlier-aware scheme that keeps precision where activations are large, and recomputing the rest. The optimizer is a mixed-integer linear program that minimizes estimated training time subject to memory constraints, and an adaptive mechanism re-solves the plan during training because activation distributions change. If the claims hold, the framework turns memory management during LLM training from a hit-or-miss rule into a continuously updated resource-allocation problem.","pith_inferences":["An implication the paper does not develop: the same per-tensor optimizer could absorb other memory levers beyond compression and recomputation, such as CPU offloading or mixed-precision storage, since it only needs a cost estimate per action.","The adaptive re-planning component suggests a natural but untested use in long-running or continually trained models, where activation statistics drift over time and a fixed plan would become stale.","A direct testable extension is to vary the per-tensor compression ratio while holding the action schedule fixed; if accuracy tracks the outliers protected by layer-specific compression, the outlier-handling claim is confirmed, and if it does not, the accuracy preservation must come from elsewhere.","Because the scheduler is formulated as a MILP, the same framework could be applied to inference-time activation memory on memory-limited accelerators, where the objective would be latency rather than training throughput."],"forward_implications":["Under a fixed GPU memory budget, the memory freed by tensor-level decisions can be reinvested in larger batches or larger models, so the reported throughput gain should compound with scale.","Static policies—recompute everything or compress everything uniformly—are dominated by the adaptive mix on models where activation sizes and recomputation costs vary across layers.","Outlier-aware layer-specific compression should allow higher compression ratios than uniform quantization before accuracy degrades, since it targets the channels that dominate reconstruction error.","As memory becomes the binding constraint, the value of the MILP planner grows with model depth and with smaller memory budgets, because the global view makes better trade-offs across layers."],"supporting_citations":[],"fun_headline_variants":["Adacc adaptively mixes recompute and compress to speed LLM training 1.37x","Adacc unifies compression and recomputation for faster, accurate LLM training","Adaptive tensor-level selections of recompute or compress speed LLM training","MILP-optimized memory plan accelerates LLM training while keeping accuracy","Adacc's adaptive memory strategy speeds LLM training up to 1.37x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the MILP's cost models are accurate: if the predicted memory footprint, compression ratio, or recomputation time for a tensor is systematically wrong, the globally optimized plan will not deliver the claimed throughput, no matter how well the optimization is posed.","fun_headline_variants_meta":{"raw":{"variants":["Adacc adaptively mixes recompute and compress to speed LLM training 1.37x","Adacc unifies compression and recomputation for faster, accurate LLM training","Adaptive tensor-level selections of recompute or compress speed LLM training","MILP-optimized memory plan accelerates LLM training while keeping accuracy","Adacc's adaptive memory strategy speeds LLM training up to 1.37x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3425,"prompt_tokens":927,"completion_tokens":2498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2392}},"tokens_in":543,"tokens_out":2498,"duration_ms":20404,"temperature":1.0,"reasoning_tokens":2392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:54:31.549756+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a transformer whose activations have no prominent outlier channels and run Adacc against a pure selective-recomputation baseline under the same memory budget; if throughput is not better and accuracy is worse, the compression branch is not pulling its weight. More directly, perturb the per-tensor memory and recomputation-time estimates fed into the MILP by 20% and compare predicted versus measured iteration time—if the plan's advantage tracks the perturbed estimates, the system is optimizing its own cost model rather than actual runtime.","supporting_citations":[],"review_version":1}