2026-10-14

Other changes reviewed today:

• In the same commit, X added a dwell-based head to the candidate retrieval model.
• Meta announced that it switched the backbone of its ads foundation model GEM from InterFormer to Kunlun and merged the per-surface models into one.
• Meta described a serving approach that uses spare inference GPU capacity to evaluate more ad candidates or route requests to a larger model.

1Video open weight went down and like and dwell weights went up in the home-mixer mirror — X

• Of the three changed weights, the like and dwell time weights went up and the video open weight went down verified — home-mixer/params/param.rs:320, 341-346, 401-406.

  • The like weight FavoriteWeight went up from 0.5 to 0.6.

    param!(FavoriteWeight, f64, "rust_home_mixer_favorite_weight", 0.6);
    

    — home-mixer/params/param.rs:320 (code), 2026-10-13

  • The video open weight VideoOpenWeight went down from 0.07 to 0.02.

  • The dwell time weight ContDwellTimeWeight went up from 0.004 to 0.006.

  • The mirror sync time is 2026-10-12T16:00:33Z.

• The VM-Ranker mirror uses the same feature switch keys, and the three values are unchanged at 0.5, 0.07 and 0.004 verified — vm-ranker/params.rs:3, 12-17, 54-59.
  • VM-Ranker is a separate service that receives the score home-mixer computed first and recomputes the value model and the DPP (determinantal point process) diversity re-ranking.
• The two mirrors diverged for the second time, after the new-author boost (PostUnexploredWeight) in the Gen4+37 commit. So again we cannot confirm which side holds the actual serving value inferred.
  • For requests that VM-Ranker answers, the VM-Ranker value may be the final one.

Two-tower retrieval builds a vector for each user and each post, and fetches the posts with the largest dot products as candidates. Which actions are trained as positives (correct answers) becomes the retrieval objective function.

• The combined preset that trains home and immersive together went from 2 candidate heads to 3 verified — phoenix/xrex/configs/xrecsys_two_tower.py:251-300.

  • The three heads are named home, immersive and ld.

  • The ld head has a single positive action, CLIENT_TWEET_RECAP_DWELLED.

            "ld_positive_actions": [recsys_pb2.ActionName.CLIENT_TWEET_RECAP_DWELLED],
            "num_candidate_heads": 3,
            "head_names": ["home", "immersive", "ld"],
            "w_ld": 0.75,
    

    — xrecsys_two_tower.py:293-296 (code), 2026-10-13

• The serving score is a weighted sum of the dot products of the home slot and the ld slot verified — recsys_two_tower_model.py:93-120, 1916-1920, 2385-2412.

  • The formula is "home slot dot product × 1.0 + ld slot dot product × 0.75".
  • Both slots sit side by side in one row of the post embedding table, and the same user vector scores each slot before the weighted sum.

• With this change, the objective of home retrieval widened from "posts to like" to "posts to like or dwell on" verified (config values).

  • Until now, the home head had a single positive, like (SERVER_TWEET_FAV).

• We infer that the head name ld is short for long dwell inferred. The evidence is that the positive is named RECAP_DWELLED and that "Recap" was Twitter's internal name for the ranked home timeline.

  • How many seconds count as "dwelled" is not in the public code.

• In the same preset, the user installed-apps feature (enable_user_installed_apps) was turned off verified (line 259).

• A model that scores up to 64 replies to a single post per user was newly released verified — xrecsys_reply_two_tower_v2.py.
  • The model name is reply_two_tower_v2.
  • The model code is recsys_reply_two_tower_v2_model.py (1,077 lines).
  • The serving registration file calls this model "reply-ranker".
• The model is much smaller than the home ranker verified. It has 1 Transformer layer, 256-dimensional embeddings and 127 history entries.
• Unlike home Phoenix, it has many hand-engineered features verified — phoenix/crates/common/xai-recsys/src/reply.rs.
  • Engagement counts of the reply
  • Author follower count and verification status
  • Whether the original post's author liked or replied to it
  • Cumulative report, block and mute counts
  • Reply score assigned by Grok (grox_score)
• The code that calls this ranker is not public verified. None of the 17 changed home-mixer files contains a call.

• For video-exclusion requests, candidates whose quoted original post is a video or whose conversation ancestor post contains a video are also filtered out verified — new filters/ancestor_video_filter.rs.
• The video carousel gained options for author dedup and SID (semantic item ID) prefix dedup, and both are off by default verified — home-mixer/util/video_carousel.rs.
  • The video carousel is the vertical video bundle in the 6th slot of For You, and it is off by default.
• One new term was added to the value model, with a default weight of 0.0 verified — xai-value-model/scoring.rs:136.
  • The term is named react_with_video.
  • The action SERVER_TWEET_REACT_WITH_VIDEO took proto ActionName slot 44, which CLIENT_CARD_CLICK had occupied.
• A retrieval dataset ZERO_FAV (No. 16), which appears to collect posts with 0 likes, and its statistics logging were added verified. No code fills the channel, so any interpretation beyond the name is inferred.
• home-mixer fills in feed request context (FeedRequestContext) values for the first time. However, the place that fills them is the ads index request, not the Phoenix ranking request verified — sources/ads_source.rs.
  • These values were added to the Phoenix request in the 2026-10-09 commit (Gen4+39).
• abuse-enforcement-service gained a new file that changes per-rule enforcement actions at runtime and a new file that turns off enforcement entirely verified. We infer it has no direct relation to feed ranking inferred. A deep dive is needed.
  • The two files are action_control.rs (3,218 lines) and off_switch.rs (1,619 lines).

• The weight mirror values changed for the first time since the Gen4+37 commit, where the two mirrors diverged on the new-author boost, and the direction is "reduce the share of video opens and increase the share of likes and dwell" verified (values). However, the VM-Ranker mirror is unchanged, so the actual serving value cannot be confirmed inferred.
• Putting dwell into the objective all the way down to retrieval is new verified. Together with the Gen4+31 commit that raised the post-click dwell weight from 0 to 0.4 and the Gen4+38 commit that changed the loss to CREAD (classification per threshold, then restoration), X appears to be increasing the share of dwell signals from retrieval through ranking inferred.
• The reply ranker is evidence that the README's claim that "all hand-engineered features were removed" applies only to the home feed verified (feature list).

↑ Contents

2GEM 2026 moved to Kunlun and merged per-surface models — Facebook·Instagram (ads)

• The backbone changed from InterFormer to Kunlun verified. Kunlun came out first in a 2026-02 paper, and its abstract already said: "Kunlun is now deployed in major Meta Ads models".

  "Compared to the previous version (Interformer), Kunlun delivers 2x performance gains with the same computation scaling." — GEM 2026 blog post (blog), 2026-10-13

• The blog post says CompSkip (a technique that turns off some components during training to skip computation) cut FLOPs by 43% verified.

  "By selectively deactivating the specific components, which is based on the assumption that the network can recover quickly after re-activating, we reduce FLOPS by 43% with negligible quality loss." — same post (blog), 2026-10-13

• Per-surface models were merged into one verified. It has shared experts and domain-specific experts together, and per-surface training data is mixed virtually rather than merged physically.

  "The 2026 generation of GEM shifts from isolated per-domain optimizations — such as Facebook, Instagram, Reels — to a single central foundation model." — same post (blog), 2026-10-13

• Long-term engagement history length was increased 2× in production, and the capability to increase it up to 25× was confirmed verified.

• Model complexity is 2.5× that of the previous GEM, and the post says it observed LLM-like scaling laws at that scale verified.

• Training efficiency (MFU 20-25%, FLOPs 4×) is the same figure as in the 2026-08-03 post verified.

• There are three paths for transferring knowledge to downstream models verified.

  • LoopFM is a post-training technique that passes GEM's internal representations and history memory to downstream models. It claims 2× the transfer effect of ordinary knowledge distillation (training that uses the teacher model's predicted probabilities as the student model's labels).

  • Soft label distillation is also used.

  • Solaris precomputes the GEM embeddings of the user-ad pairs that the next request will need.

    "It speculatively precomputes foundation model embeddings for the user-ad pairs a future request is likely to need, and delivers them just-in-time on the serving hot path." — same post (blog), 2026-10-13

• The post gave improvement figures for conversion and ad click-through rate verified.

  "GEM has delivered 18.9% conversion improvement and 12% ad click-through rate (CTR) improvement across Meta's ad surfaces and objectives, and additionally contributed 5.8%, 4.1%, and 1.6% conversion uplift on Facebook, Instagram, and off-site conversion for targeted users over the past year." — same post (blog), 2026-10-13

  The first two figures have no period or baseline, and the last three are cumulative over the past year. The A/B design is not in the text.

• It disclosed a plan to expand to organic feed data verified (plan statement).

  "Looking forward, we are excited to extend GEM to learn from all of Meta's platform data, capturing behavior across organic and ads engagements spanning text, images, and video." — same post (blog), 2026-10-13

• ⚠ Claim changed: the canon's "GEM = unification of Wukong and InterFormer" is now a description of the 2025-11 generation, and the backbone of the 2026 generation is Kunlun verified.
• Solaris points in the same direction as the design disclosed by LLaTTE (Meta's ads ranking sequence model), which moves heavy sequence computation out into an asynchronous upstream model inferred. However, the two posts do not cite each other, so we cannot tell whether they are the same system.
• For the observation these notes have tracked, that "the organic (HSTU) lineage and the ads (GEM) lineage do not cite each other", a plan to converge appeared for the first time inferred. It is only a plan statement, not a deployment fact.

↑ Contents

3Spare inference capacity adds ad candidates or larger models — Facebook·Instagram (ads)

• There are two levers for using spare capacity verified.
  • Increase the number of ad candidates evaluated in the ranking stage (Rank More Ads).
  • Send requests to a larger model (Serve Better Models). Higher-value requests are prioritized first.
• To protect the primary service, it uses physical and logical isolation, latency-based feedback and value-based load shedding (dropping less important requests first under overload) verified.
• It did not disclose effect figures for latency, conversion or revenue verified (the text has no figures).

"Inference fleets must maintain enough capacity to meet demand under strict latency and reliability targets. However, demand and capacity rarely align perfectly across time, regions, resource combinations, and accelerator types." — Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations (blog), 2026-10-12

↑ Contents

Radar

• Added the GEM 2026 generation as additional evidence to the "recommendation model scaling laws" thread (a canon update item). The evidence is the blog post and Kunlun arXiv 2602.10016.

↑ Contents

What to check next

↑ Contents

Source check notes

• X: Since 2026-10-12 there has been 1 commit, on 2026-10-13, and no release.
  • The commit is e1a0375.
• Meta: engineering.fb.com has 4 posts since 2026-10-12.
  • Two posts besides the 2 above were skipped as unrelated to feeds. One is "OCP Summit 2026" on network hardware, and the other is "Shipping Media Over QUIC at Meta Scale for Instagram Live Ingest" on live streaming.
  • The latest post on ai.meta.com/blog is from 2026-07, so there are no new posts.
  • The update stamp on the transparency.meta.com ranking explanation pages is still 2023-12-31.
  • WebSearch cross-checking returned only Threads Dear Algo and Your Algo, which are already in the canon.
• Netflix: The latest post in the netflixtechblog.com RSS is from 2026-09-25, so there are no new posts. The latest entry in the research.netflix.com publications archive is 2026-02-24.
• arXiv cs.IR: 30 items were parsed, and 0 passed.
  • The near miss is 2610.10920. It is a RecSys 2026 OARS workshop paper that evaluates the placement of recommendation modules on e-commerce search pages offline with an LM; it is limited to the search surface and makes no deployment claim.
  • The latest submission date in this response is 2026-10-09, and the 10-12 batch largely overlaps with this list.
• It was a Wednesday, so the industry blog scan was not done.

↑ Contents