Updated 21 October 2026. In May we described four dwell heads, cited phoenix/run_pipeline.py and weighted_scorer.rs, and leaned on a 2023 dwell weight of 0.0001 per millisecond. Those files no longer exist, and on 13 August 2026 the repo published the real default weights. This update covers the dwell family as the code has it now: dwell 0.05, dwell time 0.006, click dwell time 0.4, not dwelled -0.02, two “seconds” signals switched off, and a 10-second video threshold whose weight is 0.
In the 2023 Twitter ranker, dwell was a single weighted term, set so low that most coverage treated it as a rounding error. The current open-source code is different. The README’s list of 25 predicted actions has a whole “Attention” group (video quality view, dwell, dwell time, click dwell time, video continuation seconds and profile visit seconds), and a “not dwelled” signal sits in the negative group. This article walks Twitter dwell time signal by signal, with the published weights, what each one actually measures, and the writing patterns that earn it.
It is a companion to the guide to all 25 signals, which covers the full weight table, and the negative-signals article, which covers the offset branch and the not-dwelled side in context.
The dwell family, side by side
All weights below are defaults from home-mixer/params/param.rs, which says it mirrors the live configuration (last sync 20 October 2026). They are combined in compute_weighted_score in xai-value-model/scoring.rs.
The dwell-related signals and their default weights, 21 October 2026.
Signal | Default weight | What it predicts | Status |
|---|---|---|---|
Dwell | 0.05 (DwellWeight) | Probability the viewer pauses on the post | Active |
Dwell time | 0.006 (ContDwellTimeWeight) | A predicted duration of the pause | Active |
Click dwell time | 0.4 (ContClickDwellTimeWeight) | A predicted value for time spent after clicking through | Active |
Not dwelled | -0.02 (NotDwelledWeight) | Probability the viewer scrolls past without pausing | Active |
Video continuation seconds | 0 (not set in param.rs) | P(video open) times predicted seconds of continued watching | Off by default |
Profile visit seconds | 0 (not set in param.rs) | P(profile click) times predicted seconds on the profile | Off by default |
Bidirectional follow dwell boost | 0.0 | Would add to the dwell weight for original posts from mutual follows | Tested in July 2026, not shipped broadly |
Source: home-mixer/params/param.rs, home-mixer/scorers/value_model.rs (the two “seconds” weights), docs/BIDIRECTIONAL_BOOST_CHANGE.md. Each weight multiplies a predicted probability or predicted value for one viewer, never a count.

Where the dwell family enters the score: the dwell-related terms in compute_weighted_score. Source: xai-org/x-algorithm, xai-value-model/scoring.rs, captured 21 October 2026.
Two things stand out. First, per unit of probability, every dwell weight is small next to a like (0.66) or a reply (5.0). Second, the active set is four signals, not six. The two “seconds” signals are wired into the formula but switched off, and the mutual-follow dwell boost was tested and left at zero.
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 compared with the chance of a reply or a share. A small weight on a large probability adds up, and a post that earns the read moves both terms in its favour at once. The code does not let us put an exact number on that, and we will not invent one.
Dwell and not dwelled: the pause, measured both ways
The two probability signals are mirror images. Dwell is the chance the viewer pauses on your post. Not dwelled is the chance they scroll straight past. In the Phoenix action list in phoenix/crates/serving/xai-recsys-proto/proto/recsys.proto, they are CLIENT_TWEET_RECAP_DWELLED (number 11) and CLIENT_TWEET_RECAP_NOT_DWELLED (number 12), and phoenix/xrex/data/recsys/constants.py maps the “recap dwelled” training label to the first.
Two corrections from May. We called 11 a verified head index from phoenix/run_pipeline.py. That file is gone, and 11 is the action’s number in the enum, not a position in the model output. We also built a section on “recap dwell versus realtime dwell” in the mobile client. Nothing in the current code supports those details, so that section is gone. What the code does not tell us is the exact pause threshold that counts as a dwell.
The weights: +0.05 for a pause, -0.02 for a scroll-past. Our read: moving a viewer from “likely to scroll past” to “likely to pause” helps on both terms. Neither is large on its own, which is why dwell works as a steady background signal rather than a jackpot.
The mutual-follow boost deserves a note. docs/BIDIRECTIONAL_BOOST_CHANGE.md says a dwell boost for mutual follows was tested in the July 2026 A/B test alongside the reply boost, but not shipped more broadly. BidirectionalFollowDwellWeightBoost is 0.0 today, while the reply boost is 15.0.
Dwell time: a predicted value, not a yes or no
Dwell time is not a probability. It is a predicted continuous value, and the README says the weights scale “predicted continuous values, e.g. dwell time” as well as probabilities. Its weight, ContDwellTimeWeight, is 0.006.
Be careful with units. The scoring code does not state them. The Phoenix code converts dwell time into seconds when it builds a viewer’s history (phoenix/crates/serving/xai-recsys-engine/src/util.rs), and the training configuration caps dwell time at 300, which suggests seconds. But the serving units are not spelled out, so we will not turn 0.006 into “worth N likes per second”.
That is also why our May math is withdrawn. We multiplied a twelve-second dwell by a 2023 weight of 0.0001 per millisecond and concluded it beat two likes. Both the weight and the method no longer apply.
One more code fact matters for creators. The Phoenix README says each position in a viewer’s history carries dwell time, alongside the post, its author and the action taken. So how long people linger on your posts is not just a target the model predicts. It is part of what the model reads about each viewer the next time it scores you. Our read: a habit of earning real reading time from your followers builds a history that works in your favour.
Click dwell time: the heaviest weight in the family
Click dwell time is about what happens after a viewer leaves the feed: the time spent once they click through. Its weight, ContClickDwellTimeWeight, is 0.4, the largest in the dwell family and larger than the click itself (0.3 for a post click, 0.2 for opening a link).
What exactly it measures in production is not fully settled by the public code. One Phoenix training configuration (phoenix/xrex/configs/xrecsys.py) sets it up as a yes-or-no target: did the viewer stay more than 10 seconds after clicking. If production uses that form, 0.4 multiplies a probability. If it uses a raw duration, it multiplies a predicted time. Either way it is a prediction, not a count. The action list also includes external link session buckets, from under 3 seconds to over 60 seconds, so the platform clearly tracks how long people stay after leaving.
Our read on clickbait: the weights reward the click a little and staying a lot more. A hooked click that bounces in two seconds earns the cheap part and misses the expensive part. In May we said click dwell could be “three orders of magnitude larger” than in-feed dwell or even negative after a bounce. The code supports neither claim, so both are gone.
[ from the team behind this blog ]
Post on LinkedIn and Twitter in a way that reads like you
VoiceMoat is the Personal Brand OS for LinkedIn and Twitter. audenAI, the brain inside it, learns from your own posts, then helps you create, schedule and grow.
Creator plan $35 a month after the trial. Cancel anytime.
Video: the 10-second line, and a weight of zero
In May we said the video quality view (VQV) head was gated by an unpublished MIN_VIDEO_DURATION_MS threshold, guessed that it sat around eight to ten seconds, and called VQV the head that decides whether a video scores at all. Here is what the code says now.
The threshold is 10 seconds.
MinVideoDurationMsis 10_000 inhome-mixer/params/param.rs, andvqv_eligibleinhome-mixer/util/candidates_util.rsonly counts videos longer than that. The same function also skips VQV for viewers with 10,000 or more followers.The weight is zero.
VqvWeightis 0.0. So at default settings, crossing the 10-second line adds nothing to the score.Video continuation is off too. The video continuation seconds term is fixed at 0.0 in
home-mixer/scorers/value_model.rs, and the prediction is only requested whenEnableHomeExcursionScoresis on, which it is not by default. A related viewer-level continuation term is also weighted 0.What is left: video open at 0.02, plus everything a text post can earn (dwell, replies, shares, copy-link saves).
Our read: for video creators, the 10-second rule is not worth chasing for its own sake right now. Make the video as long as it needs to be to hold attention and earn replies and shares, which is where the active weights are. If the zeros change, the threshold is already wired in, so it is worth checking home-mixer/params/param.rs from time to time.
Profile visit seconds works the same way as video continuation: P(profile click) times predicted seconds on the profile, with weight 0, and the profile click weight itself is also 0 by default.
Dwell and likes are not the same signal
A common assumption is that posts with lots of likes also earn long dwell, and vice versa. Our read is that they are correlated, but loosely. Likes respond partly to social proof. Dwell responds to whether the viewer actually reads. The same post can earn one without the other.
Quick jokes lean like-heavy because the value is delivered in the first line: the viewer reacts and moves on. Operational threads and longer posts invert that: viewers pause to read, but do not always come back to like a post they have already moved past. Generic AI-shaped output tends to do poorly on both, neither fast enough to earn the like nor substantive enough to earn the read.
For a creator optimising the score rather than visible counts, the takeaway is that some of your strongest posts may have modest like counts. The score adds the dwell terms to the like term and to the heavier signals a good read tends to trigger, such as replies and copy-link shares.
Writing patterns that earn dwell
Dwell is the part of the score most shaped by editorial choice and least by reach or social proof. A post can earn a real pause with zero existing engagement if the writing carries the eye. Three patterns tend to produce it across creator categories. This is our experience-based read, not something the code states.
The second-line payoff. A first line that opens a loop the second line closes: “I shipped three products last year. The one that worked did exactly the opposite of what advice said.” The viewer has to read the second line to get the value. Posts that deliver everything in line one lose the pause.
Specific over abstract. Names, numbers, dates, places and products make the viewer slow down to parse them. “In March 2024, three enterprise customers made up 60% of revenue” earns more reading time than “we had some big customers”. Specificity is dwell density per word.
Sequence with a reveal. Lists of three to five items where a later item breaks the pattern set by the first two hold attention through the surprise.
None of these are hooks or engagement bait. A hook that promises a payoff the post never delivers earns a click or a glance and then a bounce, which, on our read of the weights, trades the cheap signals for the expensive ones.
Two things follow for post length. Length alone does not produce dwell: a longer post earns more reading time only when it is dense, specific and pays off past the first line. And thin writing gets more expensive as it gets longer, because every extra line is another chance for the viewer to scroll on. Threads compound this, since each post in the thread is scored on its own.
Why generic AI output underperforms on dwell
This section is our read, not a code fact. A general-purpose writing model produces a recognisable register: helpful, structured, broadly applicable. That register is fluent enough that viewers register the post, and generic enough that they recognise the pattern and move on. It rarely contains a second-line payoff, a surprise in a sequence or the creator-specific detail that makes someone stop, because those are exactly the properties a model aiming for “broadly applicable” avoids.
The likely result: posts that collect a normal number of likes but under-earn on the pause and the read, and on the replies and shares that a good read tends to trigger. Creators see what looks like normal like performance and reach loss they cannot explain.
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 your content read like you, specific details and all, which on our read is what earns the pause.
What to take away
Three practical moves come out of the dwell code.
Write the second line first. If the post’s value is fully present in the opening line, there is no reason to pause. Check each draft by asking whether the second line earns the read. If not, restructure.
Watch what happens after the click, not just the click. Link opens are weighted 0.2 and click dwell time 0.4. Twitter shows you link clicks. It does not show you click dwell. Our read on a proxy: if a post gets plenty of clicks but few replies, quotes or shares, readers may be bouncing, so put the value in the post and use the link for depth.
For video, do not chase the 10-second line for its own sake. The threshold exists, but the VQV weight is 0 by default. Hold attention and earn replies and shares, which is where the active weights are.
Dwell is the part of the ranker you can most directly affect through writing alone. Voice consistency (your voice is an embedding) shapes what each viewer’s history says about you, and dwell-aware writing earns the read inside that. Together they are the working form of the “voice as a moat” thesis that the opening article of this series started with.
What is dwell time in the Twitter algorithm?
Dwell is the model's prediction of whether a viewer pauses on your post, and dwell time is its prediction of how long that pause lasts. Both feed the weighted score that ranks the For You feed, alongside click dwell time (time spent after clicking through) and not dwelled (scrolling straight past).
How much does dwell time count in the Twitter algorithm?
The default weights in home-mixer/params/param.rs are dwell 0.05, dwell time 0.006, click dwell time 0.4 and not dwelled -0.02. Each multiplies a predicted probability or predicted value for one viewer, not a count. Per unit of probability they are small next to a like (0.66), but pauses are common, so they add up.
Does the Twitter algorithm penalise scrolling past a post?
Slightly. Not dwelled has a weight of -0.02 on the predicted probability that a viewer scrolls past without pausing. It is the lightest of the five negative signals, far smaller than not interested (-52.272) or mute (-58.8).
Do Twitter videos need to be longer than 10 seconds?
The code only counts a video quality view for videos longer than 10 seconds (MinVideoDurationMs 10_000), but the video quality view weight is 0 by default, and the video continuation term is switched off too. So crossing 10 seconds adds nothing to the score at default settings.
What writing patterns increase dwell time on Twitter?
In our experience three patterns hold attention: a second-line payoff where the first line opens a loop and the second closes it, specific details instead of abstractions, and a list where a later item breaks the pattern. None of them are hooks without payoff, which tend to earn a glance and a bounce.
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
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