ByteDance Uncovers Hidden Blind Spots in DeepSeek-V4: Code Unchanged, Answers Periodically Malfunction
51 minutes ago
Beating AI Express (Dongcha Insight) reports that ByteDance’s Seed team has uncovered a consistent flaw in DeepSeek-V4’s long text processing capabilities. Researchers tasked the model with completing a code snippet; the code itself remained unchanged, only the preceding irrelevant comments were extended, and the model’s answers alternated between correct and incorrect. For the base V4-Flash version, this pattern repeats every 4 tokens, while V4.1-Flash shows a similar phenomenon with a 2-token cycle. The issue is tied to DeepSeek’s memory-saving "chunked KV cache compression" method. The model merges and stores information from consecutive tokens to cut long text processing costs, but this compression results in uneven information retention across different positions, making some content harder to retrieve accurately later. ByteDance labels this periodic retrieval discrepancy "phase sensitivity." In a retrieval test with roughly 128,000 tokens, the base V4-Flash version had a maximum accuracy gap of 40.2 percentage points across different positions. Post-training narrowed this gap to 19.1 percentage points; V4.1-Flash further reduced it to 6.1 percentage points, though the periodic fluctuation persisted. ByteDance also trained multiple control models from scratch, finding the fluctuation cycle matches the compression stride, while control models without chunked compression showed no such pattern.
JPYC submits information to the Japan Fair Trade Commission regarding its JPY stablecoin payment incentives.
9 minutes ago
SoftBank seeks to raise up to $100 billion to ramp up large-scale artificial intelligence investments.
9 minutes ago
Justin Sun Outlines Five Key Opportunity Sectors in the AI Era: Crypto, Academic Data, Biopharmaceuticals, Finance and Insurance, and AI Basic Services
9 minutes ago
Sector pullback leads storage giant to exit at break-even, closing over $63.7 million in long positions in NVIDIA and storage assets.
9 minutes ago
Nasdaq CEO: Tokenization can unlock billions of dollars in trapped capital, AI is critical for 24/7 trading.
9 minutes ago
DWF Labs subsidiary sues BitGo for breaching over-the-counter trading agreement, seeking $141 million in damages.
9 minutes ago
Hot feeds
A trader profits $448K by monitoring #Binance's new listings!
2024.12.13 17:37:29
Last week, funds have flowed into #Bitcoin, #Ethereum, and #Hyperliquid.
2024.12.16 14:48:36
A $PEPE whale that had been dormant for 600 days transferred all 2.1T $PEPE($52M) to a new address.
2024.12.14 10:35:27
When Elon Musk tweeted about Moltbook, the meme coin MOLT experienced a short-term 30% price surge, hitting a new all-time high of $114 million.
2026.01.31 18:37:29
A smart #AI coin trader made $17.6M on $GOAT, $ai16z, $Fartcoin,$arc.
2025.01.05 16:05:18
A sniper earned 2,277 $ETH ($8.3M) trading $SHIRO within 18 hours!
2024.12.03 23:09:08
MoreHot Articles

How did I turn $1,000 into $30,000 with smart money?
2024.12.09

10 promising AI Agent cryptos
2024.12.05

The 30-Year-Old Entrepreneur Behind Virtual, a Multi-Million Dollar AI Agent Society
2025.01.22

10 smart traders specializing in MEMEcoin trading on Solana
2024.12.09

A trader lost $73.9K trading memecoins in just 3 minutes — a lesson for us all!
2024.12.13

What is $SPORE? Let us take you through the on-chain records to show you how it works.
2024.12.25

