Today for AI
HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 10/9/2026

ByteDance Seed Reveals DeepSeek-V4 Long-Context Drift: Chunked KV Cache Compression Causes Phase Sensitivity with 40% Accuracy Fluctuation

2 reports archived0 independent sourcesupdated 10/9/2026, 15:11:55
Synthesis & Latest Updates

ByteDance's Seed team published a paper on arXiv revealing significant performance drift in DeepSeek-V4 during long-context tasks. The study identifies that the model's chunked KV cache compression, while reducing memory costs, introduces "phase sensitivity," causing systematic asymmetry in retrieval difficulty based on token position, with accuracy fluctuations reaching up to 40 percentage points.

Cross-verified sources indicate that DeepSeek-V4 exhibits "dual nature" behavior in 128K context tests; merely altering the number of irrelevant filler tokens (with a period of 4) causes answers to oscillate between correct and incorrect, with a maximum accuracy gap of 40.2%. This finding exposes how average-based benchmarks mask periodic weaknesses, urging independent evaluation across different compression phases.

This event significantly impacts AI infrastructure optimization, warning developers to balance efficiency gains from compression against precision loss. It drives a shift from single-score metrics to multi-dimensional phase-aware assessments, prompting LLM vendors to re-evaluate the robustness of long-context mechanisms for future high-performance architectures.

LATEST/The industry is accelerating the development of specialized benchmarks for phase sensitivity and exploring hybrid compression strategies to mitigate long-context fluctuations in models like DeepSeek-V4.

HEAT TRENDHourly heat curve

151 fully observed hours
Now
0.0
Peak
21.916:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 2 reports · Latest first
  1. 量子位官网T2·78 pts
    • Chunked KV Cache compression in DeepSeek-V4 makes model performance highly sensitive to absolute input position, exhibiting periodic fluctuations every 4 tokens.
    • In 128K long-context retrieval tests, varying only the number of filler characters caused a maximum accuracy gap of 40.2 percentage points for the same information.
    • Traditional aggregate average scores may mask such 'phase sensitivity' issues, necessitating independent evaluation dimensions across different compression phases.
  2. IT之家 · 智能时代T2·78 pts
    • Chunked KV cache compression reduces memory costs but introduces token phase as a new positional coordinate, leading to periodic retrieval performance fluctuations.
    • DeepSeek-V4 series models exhibit significant phase sensitivity in long-context tasks, with retrieval accuracy differences reaching up to 40% across different phases.
    • Average benchmark scores may mask systematic weaknesses at specific input positions, necessitating attention to fine-grained performance distributions.