[

[

x algorithm series

]

]

Twitter algorithm negative signals: what a mute, block or report really costs

·

·

13 min read

·

·

By Prateek Singh, founder of VoiceMoat

By Prateek Singh, founder of VoiceMoat

Summarize with AIClaudeChatGPTPerplexityShareShare

Updated 21 October 2026. In May the negative weights were not public, so this article leaned on leaked 2023 numbers and claimed that one mute outweighs 50 likes. The repo published the real defaults on 13 August 2026, and on 14 August it added notes explaining that weights multiply predicted probabilities, not counts. This update uses the real weights, explains the offset branch exactly as the code runs it, retires the 50-likes claim, and adds the visibility filtering and labeling systems published in August, plus the Under the Hood tool.

Algorithm advice on Twitter is almost entirely additive: do this, post then, earn that signal. Read the scoring code and there is a second half. The Twitter algorithm also predicts five negative actions for every post, and it gives them the largest weights in the file. This article walks the negative signals: the five weights, what they actually multiply, the offset branch, the filters that act after a real mute or block, and the label systems that sit beside ranking.

It builds on the overview of the 2026 Twitter algorithm and the guide to all 25 signals, which cover the architecture and the positive weights.

The five negative signals and their real weights

The defaults sit in home-mixer/params/param.rs, whose first line says it mirrors the live configuration (last sync 20 October 2026).

The five negative signals and their default weights, 21 October 2026.

Signal

What the viewer does

Default weight

What else happens after a real action

Not interested

Taps “Not interested in this post”

-52.272

No filter. It becomes part of the viewer’s history

Mute author

Mutes the author

-58.8

The author’s posts are filtered out of that viewer’s feed

Block author

Blocks the author

-31.2

Filtered out in both directions. Also counted by agatha’s blocks-per-favorite label

Report

Reports the post

-234.0

Counted by agatha’s reports-per-favorite labels

Not dwelled

Scrolls past without pausing

-0.02

Nothing. It is the most common negative and the lightest

Source: home-mixer/params/param.rs, home-mixer/filters/author_socialgraph_filter.rs, agatha/scalding/labels/. Each weight multiplies one viewer’s predicted probability of the action.

Lines of home-mixer/params/param.rs showing NotInterestedWeight -52.272, BlockAuthorWeight -31.2, MuteAuthorWeight -58.8, ReportWeight -234.0 and NotDwelledWeight -0.02

The negative weights in param.rs. Source: xai-org/x-algorithm, home-mixer/params/param.rs, captured 21 October 2026.

The May version of this article used 2023 leaked numbers (mute and block near -100, not interested -74) as stand-ins. Those are gone. The real defaults are smaller for mute and block, and the report weight, which the 2023 leaks did not show, is the largest number in the whole file.

Why a big weight is not a big penalty

These numbers are easy to misread, and the repo goes out of its way to stop that. The comments in xai-value-model/scoring.rs (and the same block in home-mixer/params/param.rs) make three points:

  • Probabilities, not counts. Each weight multiplies the predicted probability of the action, and the weights “do not multiply raw engagement counts”. The README calls “1 report cancels out 468 likes” a misconception.

  • Rare actions need big weights to register. The weights reflect typical propensities as well as value, and a report’s baseline probability is more than 1000 times lower than a like’s. The report weight is large so that the prediction can affect the ranking at all.

  • Mass reporting is weaker than people fear. The prediction is your own likelihood of reporting, and recommendations are personalised, so reports from bad actors mainly affect recommendations for users similar to those bad actors. And an action only counts for ranking if it happens on a post served in the Home timeline. Navigating straight to a post, for example from a group chat, has no ranking impact.

Simple arithmetic on the repo’s own numbers shows why. The report weight is about 355 times the like weight (234 divided by 0.66; the README’s 468 used the July like weight of 0.5). If a report is at least 1000 times less likely than a like, the report term for a typical post is at most about a third of the like term. A big weight on a tiny probability is a small number.

That is also why our old headline does not hold. “One mute outweighs 50 likes” reads weights as counts. The mute weight is about 89 times the like weight per unit of probability, but what lands in the score is that weight times this viewer’s predicted chance of muting, which for most posts is tiny.

Our read, with that caveat: negatives matter when the model has a reason to expect them from a particular viewer. A post that tells a follower “this is not what you came for” can lift their predicted not-interested or mute probability, and at -52 and -59 even a small lift shows up in the sum.

The offset branch, exactly as the code runs it

After the weighted sum, every score passes through offset_score in xai-value-model/scoring.rs (lines 161 to 170). The constant NEGATIVE_SCORES_OFFSET is 0.001, set in xai-value-model/weights.rs. There are three cases:

  • Sum of 0 or more: final score = sum + 0.001.

  • Sum below 0: final score = (sum + N) / T × 0.001. N is the five negative weights added up as positive numbers: 52.272 + 31.2 + 58.8 + 234.0 + 0.02 = 376.292. T is N plus the positive probability weights (43.345 at defaults), so 419.637. The continuous weights, such as dwell time, are left out of both N and T.

  • All weights zero: a guard case that returns the sum, floored at 0. It does not apply with the published weights.

What that does, in plain terms:

  • It keeps the order. Both branches rise as the sum rises, so a less negative post still beats a more negative one.

  • It draws a line at zero. Every net-positive post scores at least 0.001. Every net-negative post lands between 0 and about 0.0009. So before the later adjustments, any post the model expects this viewer to net out negative on ranks below every post it expects them to net out positive on.

  • It keeps scores positive. The adjustments that follow (author diversity, the out-of-network discount) multiply the score by factors below 1. Multiplying a negative number by 0.75 would raise it. Our read: the offset exists so those discounts always push down. The code supports that read, since its alternative “multiplier before offset” path only applies the multipliers to non-negative sums.

In May we described this branch as a separate, non-linear penalty regime. That overstated it. It is a linear squeeze into a narrow band below the 0.001 line, not an extra punishment. The real effect is the line itself: crossing from net-positive to net-negative for a viewer moves a post below everything the model expects that viewer to welcome.

To see the arithmetic, take made-up numbers (not real predictions). Say the model gives a post a 2% chance of a like (0.66 × 0.02 = 0.0132) and a 0.05% chance of a mute (-58.8 × 0.0005 = -0.0294) for one viewer. Those two terms alone net to about -0.016. Whether the post ends up net-negative depends on every other term: replies, shares, dwell. The point is not the numbers. It is that a negative prediction only bites when the model actually expects it for that viewer.

Not dwelled: the lightest weight on the most common signal

In May we said the not-dwelled score lived only in a side scoring path and not in the main weighted sum. That is no longer true. not_dwelled_score is one of the 29 terms in compute_weighted_score, with NotDwelledWeight at -0.02. The underlying action is CLIENT_TWEET_RECAP_NOT_DWELLED in phoenix/crates/serving/xai-recsys-proto/proto/recsys.proto: the viewer scrolled past without pausing.

We also called it “the silent kingmaker of the negative side”. The real weight does not support that. At -0.02 it is smaller than the +0.05 for a pause. Our read: it is a gentle, constant nudge. Nearly every impression either earns a pause or does not, so it shows up everywhere, but it moves the score far less than one real not-interested prediction. The dwell article covers the attention side in full.

The filters that act after a real mute or block

The five weights handle the probabilistic case: this viewer would probably want less of this post. A separate set of filters handles the case where the viewer has already said so. These are yes-or-no rules, not predictions.

The “less of this” filters in the For You pipeline.

Filter

File

What it removes

Stage

ViewerMutedKeywordFilter

home-mixer/filters/viewer_muted_keyword_filter.rs

Posts matching the viewer’s muted keywords

Before scoring

AuthorSocialgraphFilter

home-mixer/filters/author_socialgraph_filter.rs

Posts from accounts the viewer blocks or mutes, authors who block the viewer, and quotes or reposts involving blocked accounts

Before scoring

TopicIdsFilter

home-mixer/filters/topic_ids_filter.rs

Posts outside requested topics, and posts in excluded topics

Before scoring

VFFilter

home-mixer/filters/vf_filter.rs

Posts visibility filtering said to drop

After selection

AncillaryVFFilter

home-mixer/filters/ancillary_vf_filter.rs

Posts whose parent, quoted or reposted post was dropped

After selection

Source: README filter tables and home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs.

Two corrections from May. The muted-keyword filter file is home-mixer/filters/viewer_muted_keyword_filter.rs, not muted_keyword_filter.rs as we wrote in May. And we described home-mixer/filters/vf_filter.rs as removing spam, violence and gore itself; in fact it only acts on the answer visibility filtering gives, covered below.

Two consequences for creators. First, once a viewer mutes or blocks you, your posts are removed from their For You before scoring, every time. There is no recovery short of the viewer reversing it. Second, a real mute is also an action in that viewer’s history, and mute author is one of Phoenix’s training labels (phoenix/xrex/data/recsys/constants.py). Our read: patterns of mutes from viewers like that one plausibly raise predicted mute probability for similar viewers, but how far that generalises is not visible in the code.

Visibility filtering and labels: separate from ranking

The README is blunt about this: ranking decides the order, and visibility filtering decides whether a post can be shown at all. Different services, different inputs, different rules. For each post and viewer, visibility-filtering/ answers allow, notice, interstitial or drop, and only drop removes the post from the feed. It runs after ranking, so it can remove a post but cannot move it up or down.

The labels it reads come from a separate labeling path published in August. The one most relevant to negative signals is agatha/, offline jobs that label an account from how others respond to its posts: blocks, reports and spam reports relative to favorites. A few details from the code:

  • The labels are ratios, such as BlocksPerFav and ReportsPerFav (agatha/scalding/labels/blocks_per_fav/BlocksPerFav.scala), and the favorites counted are out-of-network favorites.

  • A “mention-filtered” variant counts blocks and mutes from people the account @-mentioned.

  • The thresholds that turn these scores into visibility labels are not in the repo.

Some labels only bite for people who do not follow you. The rule list for For You recommendations in visibility-filtering/rules/levels.rs includes account labels such as “do not amplify” and “abusive high recall” that drop a post for non-followers while followers can still see it.

Our read: this is where blocks and reports can matter beyond one viewer’s feed, and it works on ratios, not raw counts. An account that earns plenty of favorites from people outside its network has a large denominator. An account that mainly earns blocks and reports does not.

Shadowban, revisited

In May we “gently retired” the shadowban, arguing that the code showed no hidden per-account flag. The August release changes that picture, so this section is rewritten.

What is still true: there is no secret moderator spreadsheet in the code, and the ranking side (the weighted score) has no per-account penalty term. What is new: account labels that limit visibility do exist, in the open, and some of them hide your posts from people who do not follow you. The README says a few things are left out to reduce gaming, such as some botmaker rules and the Grox prompts.

The labels are also no longer hidden from you. The Under the Hood tool, launched alongside the August release, shows aggregate statistics about the visibility-limiting labels on your account and posts, what each means and its effect on visibility. Since 18 September 2026 it also shows whether any posts were withheld in a country because of a legal demand, and which country. The code that builds the report is in under-the-hood/.

So the useful question is no longer “am I shadowbanned?” It is “does Under the Hood show a label, and if not, what is my predicted mix of actions doing?” Search filtering is a further documented layer. We cover it in what a Twitter search ban is, and the free shadowban checker tests that single layer rather than guessing at the rest.

The voice-drift to mute pipeline (our read)

This section connects voice drift to negative signals. Each step says what is code and what is our interpretation.

Step one: your audience followed a specific voice. Code fact: each viewer’s history, the main input to Phoenix, records the posts and authors they engaged with, the action taken and dwell time (phoenix/README.md). Our read: for your followers, that history holds a pattern of how they respond to you.

Step two: you ship a post that does not fit. It might be a new tone, a new format, a topic outside your usual ground, or template-shaped copy where your audience expects your own words.

Step three: the post is scored on its own. Code fact: candidates cannot attend to each other during ranking, so a post’s score does not depend on the other posts in the batch. Our read: an off-pattern post gets no help from your stronger posts and is judged against each viewer’s history.

Step four: the predicted mix shifts. Our read: positive predictions (replies, shares, dwell) soften, and for some followers predicted not-interested or mute rises. With weights of -52 and -59, a modest rise can push that viewer’s sum below zero and below the offset line.

Step five: real mutes follow, and they last. Code fact: a real mute removes your posts from that viewer’s feed through AuthorSocialgraphFilter, and it becomes training data. Our read: repeated drift compounds, because each real mute shrinks your reachable audience for good.

The question is not “did this one post survive?” but “what did this post do to how my followers expect to react to me?” We cannot measure that from outside, so treat the pipeline as a reasoned reading of the architecture, not a measured effect. Your voice is an embedding covers the history mechanism in more depth.

What this means for tools that do not check voice

Three kinds of AI tools for Twitter skip this problem by design. Template generators fit your post to a category template, which reads fluent but off-pattern for any one creator. Volume-first schedulers ship whatever you give them with no check in between. General chat assistants, prompted with light context, drift toward a helpful-assistant register that audiences spot within a few posts.

All three have real uses. None of them protects against the negative signals described above. That needs a check between draft and publish that compares the draft with your own posts. The tools comparison in this series looks at how several products handle it.

How audenAI checks drafts

audenAI, the brain inside VoiceMoat, learns from your own posts. VoiceMoat checks drafts in Studio for Style, how many of your usual habits a draft matches, and Stance, whether each line fits the claims and stories in your own posts. It does not change or game the ranking. It helps keep what you publish recognisably yours, which on our read of the code is what keeps negative predictions low.

What to do with this

Three practical moves come out of the negative-signal code.

Use voice as the gate, not the goal. A voice check does not tell you a draft is good. It tells you whether the draft sits inside what your audience followed you for. Inside is the necessary condition. Whether the post is sharp and useful is a separate question your editorial judgment owns.

Watch reach on comparable posts, not single bad days. Mute counts are not shown in Twitter’s analytics. The proxy is reach on similar, on-topic posts over several weeks. A steady decline there is the lagging sign that followers are reacting to you differently.

If reach drops suddenly, check Under the Hood before guessing. A sudden drop for non-followers may be a visibility label, which the tool now shows. A slow drop with no label is more likely the predicted mix of actions moving, and the fix is the content, not a workaround.

Next in this series: your voice is an embedding, on how history drives predictions, and the dwell article, which returns to the positive side of the ledger.

Frequently asked questions

Frequently asked questions

What are the negative signal weights in the Twitter algorithm?

The defaults in home-mixer/params/param.rs are not interested -52.272, mute author -58.8, block author -31.2, report -234.0 and not dwelled -0.02. Each weight multiplies one viewer's predicted probability of that action, and the results are added to the positive terms in one weighted sum.

How much does one mute hurt your reach on Twitter?

Less than the weight suggests. The mute weight is about 89 times the like weight, but it multiplies this viewer's predicted chance of muting, which is usually tiny, not a count. A real mute does have a lasting effect: that viewer's For You filters out your posts before scoring until they unmute you.

Can mass reporting or mass blocking kill a post's reach?

Not in the way people fear. The repo's code comments say ranking uses each viewer's own predicted likelihood of reporting, recommendations are personalised, and actions only count on posts served in the Home timeline, so coordinated reports via direct links have no ranking impact. Separately, agatha labels accounts from blocks and reports relative to out-of-network favorites; those thresholds are not public.

What does the offset_score branch do?

If the weighted sum is 0 or more, it adds 0.001. If the sum is negative, it computes (sum + 376.292) / 419.637 x 0.001, which places the post between 0 and about 0.0009. Order is kept, but every post the model expects a viewer to net out negative on ranks below every net-positive post, before the later adjustments.

Is shadowbanning real on Twitter in 2026?

Not as a secret flag in the ranking code, but account labels that limit visibility do exist and some hide posts from non-followers. Since August 2026 the Under the Hood tool shows the visibility-limiting labels on your account and posts, and since 18 September 2026 any legal withholding by country.

AI disclosure

AI disclosure

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'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

·

·

The 25 signals that rank a post on Twitter: reply, repost, like, stopped to read, share via copy link, mute, report and not interested, plus 17 more

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

·

·

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

·

·

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