复现结果

The paper's Table 2 Citeseer claim states that on the public split SGC reaches 71.9% mean test accuracy, about one point above the authors' GCN result, alongside all listed baselines. One measurement of this composite claim was executed: the 100-epoch, public-split SGC arm gave 71.2% versus a reported 71.9% (0.7 percentage points lower). The 150-epoch sensitivity arm averaged 71.85% but was not substituted. The GCN comparison and all other Table 2 baselines remain unassessed, so the overall claim is inconclusive. Separately, the citation setup claim is supported: 100 epochs, learning rate 0.2, and 60 hyperopt iterations were each reproduced exactly. The 15 unassessed claims, including claim-citation-setup's sibling claims, are not run failures.

The paper introduces SGVB and AEVB for variational inference with per-datapoint continuous latent variables, and reports experiments on MNIST and Frey Face, including Figure 3 marginal-likelihood estimates using the first 1,000 datapoints, 50 posterior samples per datapoint, 4 HMC leapfrog steps, and an MCEM baseline with 10 leapfrog steps, 5 weight updates, and a 90% acceptance target. Our reconstruction matched these operation counts: 1,000 datapoints, 50 samples, 4 steps, 10 steps, 5 updates, and 90% acceptance all matched. However, the reported estimator variance below 1 could not be comparably reproduced: observed values were 866.562 (Frey) and 0.033282 (MNIST) under a different variance definition, so the claim remains inconclusive.

The paper reports an MCP census from a 2026-07-27 official-registry snapshot, retaining the latest version of each distinct server, querying tools/list anonymously on remote targets without tool calls, yielding 98,291 tools. Our offline reanalysis reproduced the artifact-derived counts: 59,625 entries, 18,688 distinct servers, 9,454 records without a remote endpoint, 9,234 captured remote-target records, 4,838 successful nonempty targets, 4,318 connection failures, 74 timeouts, 4 no-tool targets, 0 HTTP auth rejections, and a median of 11 tools per responding server. The overall claim remains inconclusive: measurement-tool-calls was unassessed, and tool-invocation absence, snapshot date, and registry completeness were not independently verified.

FlashAttention-2 claims H100 80GB SXM5 forward-plus-backward attention throughput in Figure 7a, up to 296 TFLOPs/s, with PyTorch OOM at 16k. The run measured selected Figure 7a cells on an NVIDIA L4 instead of H100, so the reported values cannot be compared: e.g. FlashAttention-2 at 16k reported 296 TFLOPs/s versus observed 58.00092361049494, and PyTorch 512 reported 62 versus 5.800733954913403. The PyTorch 16k OOM claim was not executed. Matched throughput measurements are inconclusive; other claims were not assessed and remain unverified.

This paper claims EGD, a gradient-whitening method, accelerates grokking across toy, parity, modular, MNIST, and transformer settings, with efficiency claims tied to exact-SVD and RSVD variants. Only the three efficiency claims were assessed, all inconclusive. For modular multiplication (c11) and addition (c10), most epoch counts reproduced closely, e.g. p=79 exact SVD 228 observed versus 221 reported and rank-128 RSVD 222 versus 233, reversing the reported ordering; p=97 exact SVD 118 versus 107. No wall-clock measurements were executed, so timing portions remain unverified. Sparse parity (c12) results are inconclusive because the rerun used squared loss while the paper describes hinge loss.

The paper claims Table 10 reports full zero-shot evaluation rows for many models, and that its SSD implementation is 2–8× faster than Mamba's fused associative scan, beats FlashAttention-2 at 2K and above, and is 6× faster at 16K. Reproduction is partial: for the speed claim, observed values were 1.104–1.840× over Mamba, crossover at 16384 tokens (reported 2048), and 1.279× at 16K (reported 6×), measured on L4 rather than A100; the assessment is inconclusive. Table 10 rows and the remaining claims are not assessed here.

The paper claims that teacher predictive entropy distinguishes procedural from knowledge-intensive domains with high ROC AUC across training stages and model sizes. Our reproduction evaluated only the 1B checkpoints on a reconstructed 240-document benchmark; the 7B and 13B measurements were omitted due to size/budget limits. For the evaluated 1B results, observed ROC AUC values (e.g., 0.8079 vs reported 0.815) were broadly consistent, but the claim remains inconclusive due to incomplete multi-model and multi-stage coverage.

The paper introduces LongPIBench and reports key claims: prevention defenses show high attack success on long contexts, detection shows extreme trade-offs, GCG variants achieve high success, scope is limited to static workflows, and heuristic attacks, especially Authority spoof, substantially raise success. Our reproduction, currently inconclusive, allowed only a directional check of heuristics on document tasks: reported 1 vs observed 1 fraction attack success rate, with limitations from approximated model, corpus, and generation preventing strict matching. Other claims remain not assessed.