Search publications
“Overall, these results indicate that KV cache compression techniques introduce only limited additional latency for the first token, with the exception of TurboQuant, whose more complex quantization pipeline leads to noticeable overhead.” · Show in paper
Benchmarking KV-Cache Optimizations across Task Quality and System Performance for Long-Context ServingTable 3 and Figure 2 support the claim for the evaluated datasets, though GovReport results for TurboQuant are omitted due to 3-day single-run execution times.
“Each section links to the code and results, says how it was checked, and says where a result is still a candidate.” · Show in paper
The software defect claims are supported by independent reproductions on pinned commits and explicit prior-art/issue-history checks.
“Figure 2 shows that configurations with stronger OOD degradation also exhibit larger drift from the pretrained model. Higher-rank QLoRA checkpoints have larger symmetric KL divergence and larger effective dense update magnitudes. The strongest forgetting regime, QLoRA r=64, also has the largest SVD intruder excess, indicating a larger change in the leading spectral structure of adapted weight matrices.” · Show in paper
Figure 2 clearly shows monotonic increases in symmetric KL, dense RMS drift, and SVD intruder excess with QLoRA rank, exceeding the FFT baseline.
“Using audit_score as the primary metric, ChatGPT performs best across strategies (4.379–4.404), followed by Gemma (4.131–4.175), Gemini (3.728–3.989), and Qwen (3.489–3.605). Strategy effects are model-dependent: Gemini performs best with direct prompting and degrades under CoT/reflection (audit drops by 0.198 and 0.261, respectively), while Qwen improves with CoT/reflection (audit +0.116/+0.105).” · Show in paper
All reported audit_score ranges and strategy-dependent deltas match the values in Table 2.
“To assess whether the evaluating LLM introduces self-preference bias, we randomly sampled 50 papers and assessed each source’s comments using GPT-5.2 with the same prompt, claims, and post-processing pipeline used by the primary Gemini evaluator.” · Show in paper
“Read as a detector of drift, it is a poor one: 230 of the 273 drift steps are not marked down. mistral-nemo-12b marks about half its drift steps down, but also about a quarter of its holds and third courses. On phi-4-mini and qwen3-8b the third course is marked down at least as often as drift. So on none of the four base models is the retrospective a reliable detector of drift.” · Show in paper
All counts and percentages match §5.12 and Figure 4.
“For task quality evaluation, we use six representative datasets: HotpotQA and 2WikiMQA for multi-document QA, Qasper and MultiFieldQA_en for single-document QA, TriviaQA for few-shot learning, and MultiNews for summarization.” · Show in paper
“Across 63 comparable chat groups, median request throughput grows +11%, with 33 groups still improving by more than 10%. In contrast, the 66 comparable agentic groups are mostly already at or past saturation by C=200: 41 remain within ±10% at C=320, while 15 lose more than 10% throughput.” · Show in paper
The aggregate counts across 63 chat and 66 agentic groups, along with the representative configurations in Table 5, support the claimed scaling divergence.
“As shown in Table 3, Verified+ achieves the best F1 score, while Verified-E achieves the highest precision. ... Table 5 demonstrates that ensembling by averaging scores improves robustness and yields the best overall F1, whereas strict consensus is beneficial when precision is prioritized.” · Show in paper
Tables 3 and 5 directly support the reported precision and F1 trade-offs across prompt and ensemble configurations.
“An LCLM consists of an encoder with a pooling method that maps token chunks to soft tokens, an adapter to project to the decoder's embedding dimension, and a decoder that consumes the soft tokens as latent context.” · Show in paper
“Table 5 shows a clear degradation as the required evidence moves deeper into the document. For all models, performance is highest when evidence appears early (pages 1–2) and drops for later bins (pages 11–20 and 21+).” · Show in paper
Table 5 supports the degradation trend across page depth for ChatGPT and Qwen (with the noted small-N caveat for Qwen at 21+); Gemini's numbers are cited in the text though omitted from Table 5.
“while transferring 3× less data per decoding step than ShadowKV and using the least total memory among the methods we profile. At 128K context, it halves peak GPU memory and cuts total memory by one third relative to full KV caching.” · Show in paper
Tables 2 and 20 directly confirm the 3.1x transfer reduction vs. ShadowKV, lowest total memory (46.01 GB), ~50% peak GPU memory reduction, and ~34% total memory reduction vs. Full KV.