X’s Recommendation Algorithm: The Sorting Secrets Behind Behavior Prediction and Weights
AI Summary · Perspective of a Serial Entrepreneur (The following content is distilled by AI; the views belong to the original author; you can skip the original article after reading this)
The open-source algorithm from X (Twitter) in 2026 shows that the recommendation feed is filtered from 3,000 candidates through “behavioral prediction × weight.” The article breaks down the four-step process of recall, prediction, scoring, and adjustment, and provides an interactive parameter experiment tool, revealing the actual game logic between negative feedback (e.g., reports -234) and positive feedback.
- User acquisition reference: Understanding that the weight of negative feedback (reports/blocking) is a hundred times that of positive interactions…
- Algorithm practice: The author diversity limit (0.5 discount coefficient) means it is difficult for a single user to dominate the leaderboard, requiring a multi-account matrix strategy
- Data tools: An interactive web page based on open-source code is provided…
In-Depth Breakdown of X’s Recommendation Algorithm: From 3,000 Candidates to Personalized Recommendations
Many entrepreneurs believe that “Recommended for You” is simply interest matching, but the open-source repository updated by X (Twitter) in August 2026 reveals a harsher logic: the sorting basis is not what you “like,” but the probability of what you will do next.
I. Core Mechanism: Behavioral Prediction Replaces Interest Tags
X’s Home Mixer system regenerates the recommendation feed every time it refreshes, rather than reusing the old list. Its core process consists of five steps:
- Candidate Recall (3,000 items): Thunder (1,200 from followed accounts), Phoenix (1,000 via vector retrieval), and SimClusters (800 via community associations) work in parallel.
- Behavioral Prediction: The model predicts probabilities for 20+ types of behaviors for each tweet, including likes (31%), replies (4%), copying links (high-weight behavior), and even reports (0.2%).
- Weighted Scoring: Score = Σ (predicted probability × weight).
II. Key Data and Counterintuitive Findings
The public weight table reveals how the platform prices various behaviors:
- Positive Weights: Copy link (+20), reply to a mutual follow (+20), follow author (+4), like (+0.5).
- Negative Weights: Report (-234), hide account (-58.8), not interested (-43.2), block (-31.2).
Counterintuitive Truth: One report (-234) does not cancel out 234 likes. This is because the baseline probability of reporting is extremely low (more than 1,000 times lower than liking); if the weight were not large enough, negative feedback would barely affect the score. A tweet is significantly downgraded only when the model predicts with high probability that you will report it.
III. Ranking Adjustment and Cold Start
After the raw score is calculated, it must go through a “cleaning” process:
- Author Diversity: The second tweet from the same author is multiplied by 0.5, with subsequent tweets decaying to 0.25, preventing screen domination.
- Stranger Discount: Content from unfollowed accounts is multiplied by 0.75, but low-exposure new authors can receive a cold-start bonus.
- Visibility Filtering: The VMRanker system, independent of ranking, determines whether content is displayed normally, shown with a warning, or removed entirely.
IV. How Can Entrepreneurs Apply This?
For entrepreneurs who rely on X for growth (traffic acquisition), understanding this algorithm offers three key takeaways:
- Prioritize “Copy Link” Behavior: Its weight (+20) is far higher than that of a like, meaning content should have high collection value (tool lists, in-depth tutorials) rather than relying solely on emotional resonance.
- Avoid the “Report” Trap: Although the weight is extremely high, the threshold for triggering it is also high. Avoid controversial political topics or违规引导 (违规引导 translates to "guiding toward violations"), or once the model predicts a rise in report probability, traffic will be cut off instantly.
- Matrix Strategy: Leveraging the 0.75 stranger discount coefficient, a single account has a clear ceiling; a multi-account matrix is an effective way to break through traffic bottlenecks.
The article provides an interactive parameter experiment web page based on open-source code, allowing users to select different behaviors and drag weights to intuitively observe changes in tweet rankings, making it the best practical tool for understanding recommendation logic.
Original Article · Sspai — Matrix: Read Original →