[

[

x algorithm series

]

]

The Twitter algorithm's 25 signals: a creator's guide to how Phoenix ranks every post

·

·

15 min read

·

·

By Prateek Singh, founder of VoiceMoat

By Prateek Singh, founder of VoiceMoat

Summarize with AIClaudeChatGPTPerplexityShareShare

Updated 21 October 2026. When this guide first ran in May, the ranking weights were not public, so we counted 19 heads from the code as it stood then and borrowed 2023 numbers as a rough guide. On 13 August 2026 the repo published the real default weights, and its README now lists 25 predicted actions in five groups. This update replaces our inferred tables and the 2023 numbers with the published defaults, fixes file paths that no longer exist, and adds the repo’s own warning that a weight multiplies a predicted probability, never a count.

None of these weights changed in the 22 to 29 October releases. Our October 2026 algorithm update covers what did change, and checks the viral "30x a like" ratios against the code.

A post can rank above another post with six times the likes. The reason is that the Twitter algorithm does not score posts on one “engagement” number. Its ranking model, Phoenix, predicts how likely this particular viewer is to take each of 25 actions on the post, and a weighted sum of those predictions decides the order. This guide is the index of those 25 signals, with the default weight the open-source code gives each one and what that weight does and does not mean for creators.

The URL of this page still says 19. That was our May count, taken from an earlier state of the repo. The README’s “Scoring and Ranking” section now lists 25, and that is the number we use throughout. The first article in this series covers the wider architecture. This one is the deep dive on scoring.

How the score is built, in five steps

The README states the formula in one line: Final Score = Σ (weight × P(action)). The weights live in home-mixer/params/param.rs, and the arithmetic lives in xai-value-model/scoring.rs. In full, a post’s score goes through these steps:

  1. Predict. Phoenix predicts a probability for each action, plus a predicted value for continuous ones such as dwell time, for this viewer and this post.

  2. Weight and sum. compute_weighted_score multiplies each prediction by its weight and adds them up. There are 29 terms in all, and 22 of them have a non-zero weight at default settings.

  3. Offset. offset_score keeps every score positive. A positive sum gets 0.001 added (NEGATIVE_SCORES_OFFSET in xai-value-model/weights.rs). A negative sum is squeezed into the band between 0 and 0.001, so a post the model expects to net out negative sits below every post it expects to net out positive.

  4. Adjust. Three adjustments follow: author diversity (each extra post from the same author is multiplied by a decaying factor, down to a floor), an out-of-network discount (posts from accounts the viewer does not follow are multiplied by a factor below 1) and a new-author boost (authors with few impressions are lifted toward a target position).

  5. Rerank. VMRanker calls the service in vm-ranker/, which the README says reorders the list with a determinantal point process, giving up a little score for less similarity between neighbouring posts. Whether that reranker is switched on in production depends on startup flags that are not in the repo.

After selection, visibility filtering can still drop a post entirely. That is a separate system from ranking. The adjustments each have their own article in this series: author diversity and the reranker, and out-of-network reach.

One warning before any numbers. The README says the weights scale predicted probabilities, and “they do not scale the raw engagement counts”. A report weight hundreds of times larger than a like weight does not mean one report cancels hundreds of likes. The comments in xai-value-model/scoring.rs add that a report’s baseline probability is more than 1000 times lower than a like’s, which is why its weight is larger: so the prediction can affect the ranking at all.

The xai-org/x-algorithm README Scoring and Ranking section listing 25 predicted actions in five groups, the final score formula, the 468 likes warning and the three adjustments

The README’s “Scoring and Ranking” section: 25 predicted actions in five groups, the score formula, the weights warning and the three adjustments. Source: github.com/xai-org/x-algorithm, captured 21 October 2026.

All 25 signals and their default weights

The README groups the 25 predicted actions into five families. The table lists each with its default weight from home-mixer/params/param.rs. The first line of that file says the defaults are mirrored from the live configuration (last sync 20 October 2026), and the README says cron scripts keep them set to the primary production values. Experiments can still change them for a slice of traffic.

The 25 predicted actions and their default weights, 21 October 2026.

Group

Signal

Default weight

Notes

Engagement

Like (favorite)

0.66

FavoriteWeight

Engagement

Reply

5.0

ReplyWeight. Plus 15.0 for original posts from accounts the viewer mutually follows

Engagement

Repost

1.0

RetweetWeight

Engagement

Quote

5.0

QuoteWeight

Engagement

Share

2.0

ShareWeight

Engagement

Share via DM

5.0

ShareViaDmWeight

Engagement

Share via copy link

20.0

ShareViaCopyLinkWeight. The largest positive weight

Clicks

Post click

0.3

ClickWeight

Clicks

Profile click

0 (zero by default)

ProfileClickWeight

Clicks

Link open

0.2

OpenLinkWeight

Clicks

Photo expand

0.05

PhotoExpandWeight

Clicks

Video open

0.02

VideoOpenWeight

Clicks

Quoted post click

0.05

QuotedClickWeight

Attention

Video quality view

0 (zero by default)

VqvWeight. Only videos longer than 10 seconds qualify

Attention

Dwell

0.05

DwellWeight

Attention

Dwell time

0.006

ContDwellTimeWeight. Multiplies a predicted value, not a probability

Attention

Click dwell time

0.4

ContClickDwellTimeWeight. Also a predicted value

Attention

Video continuation seconds

0 (not set in param.rs)

Fixed at 0.0 in home-mixer’s scorer. Multiplied by P(video open)

Attention

Profile visit seconds

0 (not set in param.rs)

Fixed at 0.0 in home-mixer’s scorer. Multiplied by P(profile click)

Author

Follow author

4.0

FollowAuthorWeight

Negative

Not interested

-52.272

NotInterestedWeight

Negative

Mute author

-58.8

MuteAuthorWeight

Negative

Block author

-31.2

BlockAuthorWeight

Negative

Report

-234.0

ReportWeight

Negative

Not dwelled

-0.02

NotDwelledWeight

Source: home-mixer/params/param.rs (weights), home-mixer/scorers/value_model.rs (the two “seconds” weights), README.md (groups). Every weight multiplies a predicted probability or predicted value for one viewer, never a count.

Lines of home-mixer/params/param.rs showing FavoriteWeight 0.66, ReplyWeight 5.0, BidirectionalFollowReplyWeightBoost 15.0, RetweetWeight 1.0, ClickWeight 0.3 and OpenLinkWeight 0.2

The positive weights in param.rs, including the 15.0 bidirectional follow reply boost. Source: xai-org/x-algorithm, home-mixer/params/param.rs, captured 21 October 2026.

Three details the table leaves out:

  • 28 outputs, not 25. PhoenixScores in xai-value-model/phoenix_scores.rs holds three more: quoted video quality view (QuotedVqvWeight, 0 by default), post unexplored (PostUnexploredWeight, 0.015, counted only for in-network posts by default) and react with video (ReactWithVideoWeight, 0 by default).

  • 29 terms, not 28. compute_weighted_score adds one more term: a viewer-level video continuation figure, averaged across the viewer’s other candidates rather than predicted for the post. Its weight is also 0 by default.

  • Zeros are real zeros. ProfileClickWeight, VqvWeight, QuotedVqvWeight and ReactWithVideoWeight are all 0 in home-mixer/params/param.rs, and the two “seconds” signals are fixed at 0.0 in home-mixer/scorers/value_model.rs. The model can still predict them, but at default settings they add nothing to the score. That leaves 22 terms that actually count. Bookmarks are predicted and logged too, but they have no weight at all.

Why did we say 19 in May? We counted the named weights and scoring fields in the May snapshot of the code. The README’s list of 25, and the 28 outputs behind it, are the ones to use now.

The four signals everyone teaches, and what the weights say

Most algorithm coverage focuses on four signals: like, repost, reply and follow. They are the visible engagement actions, the counts on a post or a follower number you can watch. Here is what the code gives them:

  • Like: 0.66.

  • Repost: 1.0.

  • Reply: 5.0, rising to 20 for an original post (not a reply or a repost) from an author the viewer follows and who follows them back. The boost is in reply_weight_for in xai-value-model/weights.rs.

  • Follow author: 4.0.

Three things follow, with the usual caveat that each weight multiplies a probability.

First, like and repost carry the smallest weights of the four. That does not make them worthless. Likes are far more common than replies, so for most posts the predicted probability of a like is much larger than that of a reply, and 0.66 times a large probability can beat 5.0 times a small one. Our read: likes set the floor of a post’s score, not its ceiling.

Second, replies count most between people who already know each other. The +15 boost only applies when viewer and author follow each other and the post is an original. docs/BIDIRECTIONAL_BOOST_CHANGE.md shows its history: an A/B test from 10 July 2026, a value of 20 rolled out on 13 July, and 15 from 24 July after feedback that people were missing World Cup discussion from accounts they did not follow.

Third, follow author only has room to fire when the viewer does not already follow you. Our read: it does most of its work on out-of-network posts, where the viewer is meeting you for the first time.

The bigger point is what the four leave out. The single largest positive weight in the file is not on this list at all.

The signals nobody teaches

These are the predicted actions with no public counter on the post, or with a counter nobody watches. Several carry far more weight than a like. Two that 2023-era coverage loved are now switched off.

Share via copy link (ShareViaCopyLinkWeight, 20.0). The largest positive weight in home-mixer/params/param.rs, about 30 times the like weight per unit of probability. The viewer copies the post’s URL to paste somewhere else: a Slack thread, a doc, an email. Copying a link is rare, and the comments in xai-value-model/scoring.rs say the weights reflect typical propensities as well as value, which helps explain the size. Our read: posts people want to keep or pass on outside Twitter are the ones that earn it.

Share via DM (5.0) and share (2.0). Sending a post to a specific person weighs the same as a reply. It encodes “this is worth someone else’s attention”. Our read: idiosyncratic, opinionated and useful posts earn it. Generic engagement-shaped posts rarely do.

Quote (5.0). Also equal to a reply. A quote is a reply with an audience, and the weights treat it that way.

Click dwell time (0.4). The predicted time a viewer spends after clicking through. It multiplies a predicted value, not a probability, so compare it with care. The dwell article covers it in detail.

Post click (0.3) and link open (0.2). In May we described “click” as a link click. In the current code these are two separate signals: ClickWeight for opening the post itself, OpenLinkWeight for opening a link inside it.

Photo expand (0.05), quoted post click (0.05) and video open (0.02). Small weights. Our read: a detail-rich image helps mostly by holding attention, which feeds the dwell signals, not through the expand weight itself.

Profile click (0 by default). In May we called this one of the most valuable signals, based on a 2023 weight of 12. Today ProfileClickWeight is 0.0. Our read: a profile visit can still matter indirectly, since it is often the step before a follow (weighted 4.0), but the click itself adds nothing at default settings.

Video quality view (0 by default). The gate is real: MinVideoDurationMs is 10_000, so only videos longer than 10 seconds qualify (vqv_eligible in home-mixer/util/candidates_util.rs). But VqvWeight is 0.0, so crossing the line adds nothing at default settings.

Our read of the pattern: the invisible signals with real weight are about passing a post on (copy link, DM, share, quote) and about time spent (dwell, click dwell). The invisible signals popular in older coverage, profile clicks and video views, are switched off by default.

The five negative signals, and why their weights look so large

The negative group has the biggest numbers in the file:

  • Not interested: -52.272

  • Mute author: -58.8

  • Block author: -31.2

  • Report: -234.0

  • Not dwelled: -0.02

They are also the easiest to misread. Three facts from the repo itself:

  • They multiply this viewer’s predicted probability of the action. The code comments call reading weight ratios as count equivalents a common misinterpretation.

  • Negative actions are rare. A report is more than 1000 times less likely than a like at baseline, so its weight is larger to let the prediction move the ranking at all.

  • The prediction is personal, and only actions on posts served in the Home timeline count. Per the comments, mass reports from bad actors mainly affect recommendations for users similar to those bad actors, and reaching a post directly (say, from a group chat link) has no ranking impact.

So “one mute cancels 89 likes” is as wrong as the “one report cancels 468 likes” line the README warns about. (That 468 is 234 divided by the July like weight of 0.5. With today’s 0.66 the ratio is about 355, and it is still a ratio of weights, not of actions.) What matters is how likely this viewer is to mute, and for most posts that probability is tiny.

Our read, with that caveat: when the model does expect a viewer to want less of you, those weights are big enough to matter. A post that drifts from what a follower came for can raise their predicted not-interested and mute probabilities, and at -52 and -59 even a small rise shows up. The embedding article covers why drift is visible to the model.

There is also a deterministic layer. Once a viewer has muted a keyword, or blocked or muted you, filters remove the matching posts before scoring: the muted-keyword filter in home-mixer/filters/viewer_muted_keyword_filter.rs, and AuthorSocialgraphFilter in home-mixer/filters/author_socialgraph_filter.rs, which checks six conditions, including blocks in either direction and blocks on quoted or reposted authors. The weighted score handles “would probably want less of this”. The filters handle “already said so”. The negative-signals article walks both layers, the offset branch and the label systems that sit beside ranking.

Dwell: small weights on a signal that fires constantly

In May we called continuous dwell time the silent kingmaker, using a 2023 weight of 0.0001 per millisecond to argue that a 12-second dwell beats two likes. That math used numbers that no longer apply, so it is gone. Here is the dwell family as published:

  • Dwell (did the viewer pause): 0.05.

  • Dwell time (a predicted duration): 0.006.

  • Click dwell time (predicted time after clicking): 0.4.

  • Not dwelled (scrolled past): -0.02.

  • Bidirectional follow dwell boost: 0.0. It was tested in July 2026 alongside the reply boost and not shipped broadly, per the boost history doc.

Per unit of probability, a pause weighs less than a like. Our read on why dwell still matters: almost every impression either earns a pause or does not, so the dwell and not-dwelled probabilities are large next to the probability of a reply or a share. Small weights on large probabilities add up, and a post that earns the read moves both terms in its favour at once. The Phoenix README also notes that each position in a viewer’s history carries dwell time, so how long people linger on your posts is part of what the model reads about them later.

Compare two posts. One collects a pile of likes and nothing else. The other earns fewer likes, but readers stop to read it, reply, and copy the link to send to a colleague. On visible counts the first looks like the winner. In the ranker the second often is, because copy link (20) and reply (5) are heavy weights and it adds to several terms at once. That is our read of the central property of this ranker: it rewards a spread of strong predicted actions, not a pile of one cheap one.

Voice as a meta-signal

Phoenix conditions every prediction on the viewer’s history. The README describes the user action sequence, the viewer’s recent engagements, as the main input to the model. The phoenix/README.md adds that each history position carries the post, its author, the action taken and dwell time. That much is code.

The rest of this section is our read. If your past posts in a viewer’s history keep earning pauses, replies and shares from that viewer, the model has a pattern to go on when your next post arrives. If your recent posts start reading like someone else’s, that pattern applies less well, and the predictions drift toward what the model would expect from an author the viewer barely knows. Creators feel this as posts going quiet. We cannot see inside the trained model, so treat this as a reasonable reading of the architecture, not a measured effect.

This is the practical case for consistency. audenAI, the brain inside VoiceMoat, learns from your own posts, and VoiceMoat checks drafts for Style and Stance so what you publish keeps reading like you. It does not change or game the ranking. Its job is to keep your content recognisably yours, which, on our read of the code, is what keeps the strong positive signals coming and the negative ones quiet.

A practical framework for writing toward multiple signals

Five moves, each tied to a published weight. None of them games anything. Each is about earning actions real readers choose to take.

One: write things people want to keep or pass on. Copy link (20), share via DM (5) and share (2) are among the heaviest positive weights. Our read: operational how-tos, sharp frameworks, original data and specific opinions are what people paste into a doc or send to a colleague. The test is concrete: would a specific person you know send this to a specific person they know? If yes, you are writing for the heaviest weights in the file.

Two: end with something replyable. Reply is 5.0, and 20 for original posts between mutual follows. Ask one question your audience has a real answer to, or land on one opinion they will want to push back on. Our read: since the prediction is conditioned on each viewer’s history, specific prompts that match how your readers already reply work better than a generic “thoughts?”, and the boost rewards writing for the people you already have two-way relationships with.

Three: earn the read. Dwell (0.05), dwell time (0.006) and not dwelled (-0.02) fire on nearly every impression. A first line that opens a loop the second line closes makes the viewer stop: “I shipped three products in 2025. Two failed. Here is what the third did differently.” Headline-only posts deliver the payload in line one and lose the pause.

Four: add a link only when the post earns the click and the read after it. Link open is 0.2 and click dwell time is 0.4. Our read: a link that gets clicked and abandoned earns the cheap part and misses the expensive part. Linkless posts give up the click signals but often win on dwell and replies. Be deliberate.

Five: do not trade your voice for visible counts. Like is the smallest of the familiar weights, and the negative weights are the largest numbers in the file. Template-shaped posts chasing likes risk the trade that matters most: a small rise in predicted not-interested or mute from the followers who came for you. That is our read, not a code fact, but it follows from the weights.

Each move rests on a weight multiplied by a probability. None of them works unless the post earns the action from real readers. If you want to compare how your recent posts are doing on the visible part, the engagement rate calculator is a quick start.

Up next in this series: the negative-signal economy covers the five negative weights, the offset branch, filters and labels. The dwell article covers the attention signals in detail. Your voice is an embedding returns to how a viewer’s history drives the predictions.

Frequently asked questions

Frequently asked questions

What are the 25 signals in the Twitter algorithm?

The README lists 25 actions Phoenix predicts for every post, in five groups. Engagement: like, reply, repost, quote, share, share via DM, share via copy link. Clicks: post, profile, link, photo expand, video open, quoted post. Attention: video quality view, dwell, dwell time, click dwell time, video continuation seconds, profile visit seconds. Author: follow author. Negative: not interested, mute, block, report, not dwelled. Each prediction is multiplied by a weight and the results are summed.

What are the real Twitter algorithm weights?

The defaults are in home-mixer/params/param.rs. The largest positive weights are share via copy link (20), reply (5, or 20 for original posts between mutual follows), quote (5), share via DM (5) and follow author (4). Like is 0.66 and repost is 1. Profile click and video quality view are 0 by default. The negative weights are not interested -52.272, mute -58.8, block -31.2, report -234 and not dwelled -0.02.

Does one report cancel out hundreds of likes?

No, and the repo says so directly. Weights multiply the predicted probability that this viewer takes the action, not raw counts. A report is more than 1000 times less likely than a like at baseline, which is why its weight is larger. The predictions are personalised, and only actions on posts served in the Home timeline count for ranking.

Why can a post with fewer likes outrank one with more?

Because a like carries one of the smaller weights (0.66). A post that also earns replies, DM shares, copy-link shares and real reading time adds to several heavier terms at once. The score sums all of them, so a spread of strong predicted actions usually beats a pile of likes alone.

Which parts of this guide are code facts and which are interpretation?

Code facts: the 25 actions, the default weights, the 28 outputs in PhoenixScores, the 29-term weighted sum, the offset_score step and the mutual-follow reply boost. Interpretation: which content earns each signal and how voice consistency affects the predictions. We label those as our read in the body.

More in the series

Twitter's new ranker, explained: what the open-source code shows, updated October 2026, with 25 predicted actions and the ranking weights now published

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

·

·

Mutes count more than likes: mute, block, report, not interested, scrolled past and muted keywords all pull a post down

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

·

·

Your voice is an embedding: your account becomes an author token, your posts are embedded with it, each is scored against the reader's history, and consistency makes a tight cluster

Twitter Algorithm

How the Twitter Phoenix algorithm sees creators: hashed authors, semantic IDs and viewer history

Phoenix, the model behind the Twitter For You feed, never reads your "voice". It reads hashed author IDs, semantic IDs that describe each post, and each viewer's last thousand or so actions. This refresh walks the production code published since August 2026 and labels clearly where our creator advice is interpretation rather than code fact.

12 min read

·

·

Dwell time is the new like: four dwell signals, stopped to read, time spent reading and time after a click help, scrolled past hurts

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

·

·

The whole series

The Twitter/X algorithm in May 2026: all ten articles

Read ›

Read ›

[

ready to start?

]

Create more.
Stay unmistakably you.
Grow from there.

3-day free 3-day trial · Cancel anytime · Your content stays private

[

ready to start?

]

Create more.
Stay unmistakably you.
Grow from there.

3-day free 3-day trial · Cancel anytime · Your content stays private

[

ready to start?

]

Create more.
Stay unmistakably you.
Grow from there.

3-day free 3-day trial · Cancel anytime · Your content stays private