Buy Crypto
Pay with
USD
USD
Buy & Sell
Hot
Supports Visa, Mastercard, SEPA & more
P2P
0 Fees
Flexible trading, zero fees
Gate Card
Use your crypto for payments worldwide
Markets
Trade
Basic
Advanced
Futures
Futures
Access hundreds of perpetual contracts
CFD
Gold
One platform for global traditional assets
Options
Hot
Trade European-style vanilla options
Unified Account
Maximize your capital efficiency
Demo Trading
Introduction to Futures Trading
Learn the basics of futures trading
Futures Events
Join events to earn rewards
Demo Trading
Use virtual funds to practice risk-free trading
U.S. Stocks
CFD
U.S. stock CFD derivatives
Futures
High leverage, 24/7 trading
Tokenized Stocks
Backed by real stock assets
GUSD
Mint GUSD for Treasury RWA yields
Earn
Launch
CandyDrop
tag
Collect candies to earn airdrops
Launchpool
Quick staking, earn potential new tokens
HODLer Airdrop
Hold GT and get massive airdrops for free
Pre-IPOs
Unlock full access to global stock IPOs
Alpha Points
Trade on-chain assets and earn airdrops
Futures Points
Earn futures points and claim airdrop rewards
Investment
Square
Square
Post, share, and explore crypto trends
Live
moments live
Live crypto market analysis
Chat
Chat with crypto traders
News
What is happening in crypto
flower_head
More
Promotions
AI
Others
TradFi
Rewards
Square
Latest
Hot
News
Profile

Research on the Disconnection Mechanism Between Tool Use Proxy Cognition and Action

MeNews
robot
Abstract generation in progress
AIMPACT message, May 17 (UTC+8), this explainability paper focuses on tool usage agents, detecting hidden states to find that models often recognize when to call tools, but actual calls fail, with a mismatch rate of 26%-54%. The issue is entirely centered on the transition from cognition to action, rather than cognition itself. Internal detection directions can be decoded, but the final token mechanism of later layers causes signal rotation, almost orthogonal to the generated action. The research aims to predict the effectiveness of intervention measures, pointing out that common attributions such as insufficient prompts or training may overlook the geometric structure of later layers, providing a reasonable explanation for the performance ceiling in tool usage prompt A/B testing. (Source: AiHot)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 6
  • Repost
  • Share
Comment
Add a comment
Add a comment
GateUser-cf218ace
· 6h ago
The discovery of geometric rotation of the later layer tokens is so crucial. Previously, everyone was focusing on tweaking prompt engineering back and forth, but it turns out the root cause lies in the misalignment of directions in the representation space.
View OriginalReply0
FloatingTeacup
· 6h ago
The conversion bottleneck from cognition to action—this framework can be applied to many AI safety issues.
View OriginalReply0
QuietRugAlarm
· 7h ago
The word "orthogonal" is used brilliantly; signals and actions are almost perpendicular, and even the strongest cognition can't break through.
View OriginalReply0
FarmingNoSleep
· 7h ago
Geometric structure > Prompt engineering, this conclusion is too important for people building agents.
View OriginalReply0
StardustUnderTheGlassDome
· 7h ago
I thought about it, and that explains why sometimes using the same tool with different wording can still call successfully—the rotation angle has changed.
View OriginalReply0
YieldBento
· 7h ago
Internal signals are decodable, but the subsequent layer is orthogonal— is this orthogonality design a bug or a feature?
View OriginalReply0

  • Pinned

Sitemap