x algorithm series
How often to post on Twitter: author diversity, DPP and your third post of the day
10 min read
Updated 21 October 2026. Since May, xAI has published the scoring code and configuration defaults behind For You. The author diversity formula we described is real, but it lives in xai-value-model/scoring.rs, not in the scorer files we cited, and its values are now visible as code defaults. This update also corrects how the DPP reranker’s theta works, and drops our claim that a third post is “rejected”: it is discounted, not removed.
Two questions come up more than any others about Twitter posting frequency. How often should I post? And why do my recent posts do worse than older ones when the content is similar? The For You code published in xai-org/x-algorithm answers part of both, and the answer is not “post more”. For the practical side, see our posting frequency guide.
Three mechanisms touch frequency directly. An author diversity multiplier discounts each extra post from the same author in one feed request. A determinantal point process (DPP) reranker in the vm-ranker service picks a varied set from the top of the list. And two dedup filters collapse repeated reposts and branches of one conversation. None of them deletes your third post. Each makes an extra post worth less in the same feed load. This piece pairs with 48 hours in For You and the series hub.
The author diversity formula
The README describes it in one line: each post after an author’s first “is multiplied by a decaying factor, down to a floor”. The code is diversity_multiplier in xai-value-model/scoring.rs:
multiplier = (1 - floor) * decay^n + floor
Here n is the number of higher-scoring posts by the same author already in the candidate pool for this request. author_pool_counts sorts every candidate by score and counts, for each post, how many posts from the same author sit above it. The top one gets n = 0 and a multiplier of exactly 1. Each later one decays toward the floor.
Three settings control it: enable_author_diversity, author_diversity_decay and author_diversity_floor (xai-value-model/weights.rs). Their defaults sit in vm-ranker/params.rs: enabled, decay 0.5, floor 0.25. The unit test in xai-value-model/scoring.rs uses the same 0.5 and 0.25.
Be careful with those numbers. They are code defaults, not values from the synced production file (home-mixer/params/param.rs), and other defaults in vm-ranker/params.rs lag the synced values: its like weight is 0.5 against the synced 0.66. So the production decay and floor are unclear from the code. In May we called them redacted. They are visible, just not confirmed.
Author diversity multiplier at the code defaults (decay 0.5, floor 0.25)
Post from the same author, in score order, in one request | Multiplier |
|---|---|
1st | 1.000 |
2nd | 0.625 |
3rd | 0.438 |
4th | 0.344 |
5th | 0.297 |
6th | 0.273 |
10th | 0.252 |
Limit | 0.250 |
Illustration only. The shape, a fast drop then a floor, comes from the formula; the exact numbers depend on values the repo does not confirm for production.

The author diversity formula and how it is applied with the out-of-network discount. Source: xai-value-model/scoring.rs, captured 21 October 2026.
Where it runs in the pipeline
Our May version placed this step in a home-mixer file called author_diversity_scorer.rs, followed by an out-of-network rescale in one called oon_scorer.rs. Neither file exists today. Here is the current order for one For You request:
Sources gather candidates: Thunder for accounts you follow, plus Phoenix retrieval, SimClusters and popular posts out of network (nine sources in
home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs, some off by default).Pre-scoring filters run, including the 48-hour AgeFilter and
RetweetDeduplicationFilter.Phoenix predicts the chance of each of the 25 actions the README lists, from likes and replies to dwell, follows, mutes and reports.
The weighted sum,
Σ (weight × predicted probability or value), thenoffset_score, incompute_weighted_score. In code that sum has 29 terms.The three adjustments: the new-author boost, then author diversity and the out-of-network discount, applied together as one multiplier by the vm-ranker service (
vm-ranker/scoring/value_model.rs).The DPP reranker in
vm-ranker/dpp.rs, if it is running.TopKScoreSelectorkeeps the top 50.Post-selection filters: visibility filtering, then
DedupConversationFilter.
The out-of-network discount multiplies posts from accounts you don’t follow, and by default also replies and reposts from accounts you do follow, by OonWeightFactor. Its code default is 0.75 in vm-ranker/params.rs, with the same production caveat. One more detail: if the vm-ranker call fails, home-mixer serves its own local scores, and that local path switches author diversity off (home-mixer/scorers/value_model.rs).
A note on weights, since they set the scores being discounted. They live in home-mixer/params/param.rs (like 0.66, reply 5.0, share via copy link 20.0, report -234.0) and they multiply predicted probabilities, not counts. The repo’s comments say reading them as “one report cancels 468 likes” is wrong, partly because the baseline chance of a report is more than 1,000 times lower than a like. Our guide to the 25 predicted actions and negative-signal guide cover them in full.
The discount is per request, not per day
The diversity count only sees one candidate pool: the posts gathered for one viewer’s one feed load. Your posts compete with each other only when several of them are in that pool together. That rules out anything older than 48 hours (AgeFilter) and anything that viewer has already been shown (the seen and served filters).
Our read on what that means in practice:
A follower who opens Twitter often meets your posts a few at a time. Each load holds fewer of your unseen posts, so less of the discount applies.
A follower who opens it once a day meets every unseen post you published since their last visit in one pool. That is where the 3rd, 4th and 5th posts are discounted hardest.
So the right cadence depends on how your audience reads, not on a universal number of posts per day.
What the formula does not do
It is easy to read more into author diversity than the code supports. Here is what it does not do:
It does not count your posts per day. It counts your posts in one viewer’s candidate pool, ranked by score. A post from yesterday that the viewer has already seen is not in the pool, so it does not count.
It does not touch your best post. The highest-scoring post from you in a request always gets a multiplier of exactly 1.
It has no memory across requests. Each feed load starts the count again. Nothing in the scoring code tracks how often an account posts over time.
It does not remove anything. Every post keeps at least the floor share of its score. Removal only happens later, if DPP skips a post or a filter drops it.
One related cap sits outside the formula. The shared popular-posts list, built from the most-followed active accounts, keeps at most five posts per author (home-mixer/util/popular_posts.rs, code default). It only matters if your account is large enough to be in that list.
DPP: the variety pass
The README describes vm-ranker as reordering posts with a determinantal point process over their embeddings, “giving up a little score for less similarity between neighbours”. In vm-ranker/dpp.rs it works like this:
It takes the top
max_selected_rankposts by score. The code default (DppMaxSelectedRank) is 150.It gives each post a quality value, exp(alpha × score ÷ top score), where alpha = theta ÷ (2 × (1 - theta)).
It measures how alike each pair of posts is, using the cosine similarity of their content embeddings. A repost uses the original post’s embedding.
It greedily picks the set that is both high quality and spread out, up to
top_kposts (50 by default invm-ranker/args.rs).Posts it picks keep their score and order. Posts it does not pick get a score of 0 (
vm-ranker/scoring/dpp_model.rs), so they drop out of the top 50.
Two corrections to May. First, we called theta “the diversity weighting”. In this code a higher theta gives score more say: at theta 0 the pick ignores score and only seeks variety, and as theta nears 1 score dominates. The code default is 0.65. Second, we said DPP runs whenever theta is above zero. It runs when DppEnabled is on (code default true) and the service was started with its DPP flag, whose production value is not in the repo. It is skipped for topic requests by default.

Code defaults for author diversity (0.5, 0.25), the out-of-network factor (0.75) and DPP (theta 0.65, pool 150). Source: vm-ranker/params.rs, captured 21 October 2026.
Our read for creators: DPP compares content, not authors. Two near-identical posts compete with each other here even if two different people wrote them, while two genuinely different posts from you can both survive DPP. Author diversity still discounts the second one.
A worked example: ten posts in a day
Assume a follower opens For You once in the evening, all ten of your posts from that day are unseen and under 48 hours old, and Phoenix scores them about the same. At the code defaults, the diversity multipliers run 1.0, 0.625, 0.438, 0.344, 0.297, 0.273, 0.262, 0.256, 0.253 and 0.251.
Add them up and ten posts carry about 4.0 posts’ worth of score in that feed load. Five posts carry about 2.7. Three carry about 2.1. Each discounted post still has to beat hundreds of other candidates for 50 places, and DPP will drop any that look too much like your others.
That is our arithmetic on code defaults, not a measured result. If production uses different values, the numbers move but the shape stays: steep at first, then flat. Thunder’s per-author caps (50 original posts and 30 replies or reposts per author per request) do not bind at ten posts a day; see 48 hours in For You.
Dedup filters: a different mechanism
Two dedup filters in home-mixer/filters/
Filter | Stage | What it does |
|---|---|---|
Before scoring | Keeps one candidate per original post across all its reposts (the first one found) and merges the sources that found it | |
After selection | Keeps only the highest-scored post per conversation, where a conversation is identified by its earliest ancestor |
In May we said the conversation filter usually kept the original or the most-scored reply. The code is simpler: the highest score wins, whether that is the opener or a reply. The repost filter is also why a viral post does not stack up in one feed when many people you follow repost it.
Threads and posting frequency
Threads matter here because each reply in a thread is its own post. In code:
Thunder returns your thread replies under the replies-and-reposts cap of 30 per author.
They sit in the same candidate pool as your other posts, so they count toward n in the author diversity formula.
After selection,
DedupConversationFilterkeeps one post per conversation, so a follower gets at most one post from your thread in a single feed load.For people who don’t follow you,
OONRetweetReplyFilterremoves out-of-network replies, so only standalone posts such as the opener travel out of network.
In May we described long threads as paying an exponential penalty for each extra post in the visible feed. That was wrong: the extra posts are mostly collapsed by the conversation filter, not shown at a discount. Our read: the thread’s best single post carries it in For You, so make the opener work on its own, and expect a long thread’s replies to count against your other posts in the same window.
So how often should you post on Twitter?
The code has no posting-frequency penalty and no ideal number. What it has is diminishing returns inside each feed load, a variety pass that punishes near-duplicates and a 48-hour window per post. Our read on how to post with that in mind:
A few strong posts beat many similar ones. At the code defaults, the third post in one load keeps under half the score of the first.
Spread posts out. Posts that land in different feed loads for a follower are not discounted against each other.
Vary the content. DPP compares meaning, so five takes on one idea compete with each other, even across authors.
Earn the actions the model predicts. Replies, dwell, shares and follows raise the score that gets discounted; nothing in the diversity step adds them back.
VoiceMoat’s point of view is that consistent, recognisable content from a real person tends to earn those positive actions and avoid the negative ones (not interested, mute, block, report). That is our interpretation, not something the code states, and no tool changes how For You ranks a post. It is why we think tools built around posting volume optimise the wrong variable. We compare approaches in VoiceMoat vs Tweet Hunter and Typefully, and the series hub covers the full ranking path.
How often should you post on Twitter according to the algorithm?
The open-source code sets no ideal number and has no posting-frequency penalty. It discounts each extra post from the same author within one feed load, so a few strong, varied posts spread across the day keep more of their score than many similar posts at once. That advice is our read of the code.
Does posting more than once a day hurt your Twitter reach?
It does not lower your first post. Extra posts that land in the same feed load are multiplied down: at the code defaults, ×0.625 for the second, ×0.4375 for the third, then toward ×0.25. Production values for these settings are not confirmed in the repo.
What is author diversity in the Twitter algorithm?
A multiplier in xai-value-model/scoring.rs: (1 - floor) × decay^n + floor, where n is how many higher-scoring posts by the same author are in the same request's pool. The code defaults in vm-ranker/params.rs are decay 0.5 and floor 0.25.
What is DPP in the Twitter algorithm?
A determinantal point process reranker in the vm-ranker service. It takes the top 150 posts by score (code default), picks a set that is high-scoring and not too similar in content, and gives the posts it skips a score of 0. Whether it runs in production depends on a startup flag that is not in the repo.
Do Twitter threads count against posting frequency?
Partly. Thread replies are separate posts, so they count toward the author diversity discount in the same feed load. But DedupConversationFilter keeps only the highest-scored post per conversation, so a follower gets at most one post from a thread in a single load.
This article was drafted with AI assistance and human-edited by Prateek Singh. All technical claims are sourced to the xai-org/x-algorithm repository; file paths are cited inline.
More in the series

Twitter Algorithm
The Twitter algorithm in 2026: how the open-source ranker works and why your voice matters
Twitter's For You feed runs on open-source code in the xai-org/x-algorithm repo on GitHub, first published in January 2026 and updated almost every weekday since 13 August. A transformer called Phoenix predicts 25 actions for each post, published weights turn them into one score, and author diversity, an out-of-network discount, a new-author boost and a DPP reranker follow. This hub walks the current code in plain English and explains why consistent, recognisable content still matters for creators.
14 min read

Twitter Algorithm
The Twitter algorithm's 25 signals: a creator's guide to how Phoenix ranks every post
Phoenix, the model behind the Twitter algorithm, predicts how likely each viewer is to take 25 different actions on a post, and a weighted sum of those predictions sets the order. This guide lists all 25 with the default weights published in home-mixer/params/param.rs in August 2026, from share via copy link at 20 to report at -234. It explains why every weight multiplies a predicted probability, not a count, and closes with a five-move framework for posts that earn several strong signals at once.
15 min read

Twitter Algorithm
Twitter algorithm negative signals: what a mute, block or report really costs
The Twitter algorithm predicts five negative actions for every post and gives them the largest weights in the code: report -234, mute -58.8, not interested -52.272, block -31.2 and not dwelled -0.02. Those weights multiply one viewer's predicted probability of the action, not counts, so a single mute does not cancel dozens of likes. This article explains the offset branch exactly as the code runs it, the filters that act after a real mute or block, and the visibility labels you can now check with Under the Hood.
13 min read

Twitter Algorithm
Twitter dwell time: the real weights behind every pause on your posts
The Twitter algorithm has a whole family of dwell signals, and since August 2026 their weights are public: dwell 0.05, dwell time 0.006, click dwell time 0.4 and not dwelled -0.02. Two more "seconds" signals and a mutual-follow dwell boost are wired in but set to zero, and the 10-second video threshold carries a weight of 0. This article explains what each signal measures, why small weights still add up, and the writing patterns that earn the pause.
11 min read

Twitter Algorithm
Twitter algorithm recency: why For You drops posts after 48 hours
How long do tweets stay in the For You feed? In the open-source code as of 21 October 2026, the answer is 48 hours: an AgeFilter removes older posts from every candidate source before scoring. There is no hand-set recency multiplier, but Thunder serves followers' posts newest first, the Phoenix training code reads post age as a feature, and the new-author boost only covers posts up to two hours old. This refresh corrects our May claim that old posts live on out of network.
10 min read
The whole series
The Twitter/X algorithm in May 2026: all ten articles