What I took away
from RecSys 2026

Exploration was my most important theme at the conference. I was especially interested in approaches that use the model's beliefs to decide what to show or what to learn from, rather than relying only on rules about which content deserves more exposure.

What could a response teach us?

EGRec was one of the papers I cared about most because it asks what a response could teach the system about the person. It adds hypothetical feedback to the history, re-encodes the user with model weights frozen, and measures the change in catalog-wide relevance using Jensen-Shannon divergence. This is an epistemic-gain signal: how much a response could revise the model's understanding, not a Bayesian estimate of uncertainty about an item's reward.

A small MLP learns these gain values offline from the frozen recommender's hidden features, avoiding hypothetical catalog-wide rescoring at serving time. I particularly liked using the predicted gain as an extra context feature for Neural Thompson sampling: the bandit supplies reward uncertainty, while EG supplies a direction for interest discovery. That combination had a strong interest-coverage result in one offline setting, but gains varied across experiments.

Put uncertainty to work inside the ranker

YouTube's Uncertainty-Aware Reward Modeling was another favorite because it puts model-based exploration directly into a production ranker. One approach samples scoring weights from a Bayesian last layer; another predicts reward percentiles and uses their spread as a scoring bonus. Neither requires making the entire large model Bayesian. Both report engagement and discovery gains, although the appropriate scoring policy depends on the device and combining the approaches does not automatically help.

What I valued in Breaking the Loop was its practical comparison of freshness rules, training changes and uncertainty-based methods. An age-based boost did not significantly improve new-release engagement; an uncertainty bonus did, but reduced artist diversity. Even an uncertainty head unused at serving improved new-release engagement, a revealing control whose cause remains unresolved. The breadth of these ablations shows which interventions worked, which did not, and the trade-offs involved in moving from familiar rules to model-based alternatives.

Training emphasis and reward design matter too

Soft Curriculum Learning makes a relatively contained training change: retain the examples, but repeatedly shift their loss weights from easier to harder examples. Difficulty can come from engagement counts or uncertainty, and the engagement-based version delivered the strongest ranking gains. I found its retrieval result especially instructive: unique retrieval engagements rose by 7.7% online despite lower offline mean average precision. That aggregate offline metric did not capture the benefit of the changed training emphasis.

With MoR, it was the reward design that interested me. It blends a long-term affinity score with samples from a Beta posterior updated by recent feedback. Without new impressions, the reward exponentially returns toward the long-term score. This allocates retrieval capacity across content channels, not retrieval models. Its long-term-only baseline captures most of the online gain, so the interesting contribution for me was the controlled adaptation, not an assumption that the bandit explains the headline result.

From Placements to Pages interested me because the decision is about how recommendation modules work together on a page, not how each slot performs alone. A Transformer conditions on the modules already chosen, and a bandit learns from feedback across requests. This is not responsiveness to interactions within the current page: the composition is chosen before those interactions happen. I found that distinction useful, along with the practical choice to limit strict sequential decisions to the first two placements to control latency.

The exploration papers made the next question sharper: what should the model retain from what it learns? I was less interested in an embedding being good in the abstract than in what it lets the next model do. Several papers changed either the information it carries or where that information enters the computation.

Where a feature enters matters

WHALE was one of the most practically relevant papers for me. Its feature-interaction branch repeatedly queries the behavior-history branch through attention, layer after layer. The features do not wait until the end to meet the sequence representation; they help determine what is read from the history throughout the model. That was the important idea for me: changing where information enters can change the computations available to the model, even before adding new information.

The paper compares models at matched compute and tests the contribution of individual fusion components, rather than only showing that a bigger model performs better. Those comparisons do not isolate every benefit of repeated fusion, and the serving-throughput trade-off still matters. The interesting question is when richer interaction between the feature and history branches justifies that cost.

Google Discover's Heterogeneous Ranking felt like a focused improvement rather than a new paradigm. Content and context signals enter the mixture-of-experts gate and generate scale-and-shift vectors for the experts' outputs. That is more specific than adding content type as another input: the context changes how expert features are expressed. I also liked the evaluation's check on ordering across formats, where improved within-format accuracy alone can miss poor comparisons.

Identity, interests and the downstream task

Space Efficient Item Embeddings keeps larger dedicated vectors for popular items, while representing the tail through hashing or low-rank projection. The point is not to shrink every vector equally: allocate capacity according to the data available for each item, preserving catalog coverage within a memory budget. I found that uneven allocation useful well beyond one recommendation architecture. It is a memory result, not evidence of a serving-latency gain.

SEMCo learns interaction-supervised content similarities instead of copying collaborative embeddings. Its sampled-entmax loss gives uninformative negatives zero gradient; the sparse-representation follow-up instead sparsifies the stored vectors through activation and top-k selection. Sparse supervision and sparse representations are distinct mechanisms. Neither paper reports a live A/B test.

On the user side, Interest Sequence builds an interest hierarchy through semantic-neighbor graphs and repeated clustering, then records lifetime and recent feedback at its nodes. Selected interests become tokens containing IDs and behavior statistics; learned global tokens gather information through a Transformer trained with the ranker. I liked the richer record of the person, not only the encoder. Constructing that record and learning how to aggregate it remain distinct contributions.

The part of Mosaic that most interested me was CoTrain, not the entire model-fleet architecture. It adds a copy of the downstream ranker during training and shares the user tower with it, letting that ranker's gradients shape the upstream embedding directly. Serving remains decoupled. That is the concrete attraction: train the representation for the model that will consume it. The broader online gains cannot be assigned to CoTrain alone.

Pinterest's onsite-offsite graph links permissioned offsite conversions and onsite engagement through shared users. TransRA learns entity embeddings in a common space; those embeddings initialize trainable ID tables that the ranking loss can refine. I liked both parts: bringing information across sources and adapting it to the recommendation task, since frozen embeddings alone gave limited gains. Richer representations still bring a practical question: what does it cost to build and use them?

My interest here was practical: how can richer features and longer histories fit within a constrained memory and serving-latency budget? These papers offered several different places to make that trade-off, rather than treating a larger Transformer as the only next step.

Token Factory was my favorite paper of the day

Token Factory reduces the input before expensive attention begins. It concatenates an event's normalized or embedded features and maps them through a jointly trained encoder, such as an MLP, into soft tokens. A further MLP along the history-length axis or attention pooling can compress events into fewer positions. These are learned floating-point vectors, not discrete vocabulary codes or lower-precision numbers: the saving comes from giving the sequence model fewer positions to process.

What made this my favorite paper of the day was the control it gives the model designer. Feature richness, retained history, compression and model capacity become choices that can be balanced within a budget. YouTube evaluates the approach in ranking and generative retrieval, with training-efficiency improvements and online gains in retrieval contribution and fresh-content exposure. The online changes are bundled, and the paper does not establish an end-to-end serving-latency gain from compression alone. Even with that limitation, this felt like a foundation for making sequential recommendation more practical.

Rate-distortion: what history is worth keeping?

What I loved about Information-Aware Long Sequence Compression for Sequential Recommendation (RDSR) was its rate-distortion framing. A learned selector keeps a fixed number of events before the main sequence model, then compresses their representations. A Lagrangian-style objective balances next-item prediction against a KL information penalty, making retained information and prediction quality an explicit trade-off. The selected length is a separate control, not the definition of information rate. The selector still reads the full history, and production latency savings remain unproven.

Attend to an abstraction, not always the raw history

SequenceO1 makes the reusable summary explicit: each history event distributes its weight across a fixed set of learned prototypes, which aggregate the events into a short sketch. Because that sketch does not depend on the candidate item, it can be cached and reused; candidate-conditioned attention reads it alongside recent behavior. What appealed to me was avoiding repeated work on older history, not merely shrinking each event vector. The paper reports online gains, but no end-to-end latency benchmark.

Two narrower ideas also caught my attention. TSD-SR learns a polynomial spectral filter over position-ordered embeddings instead of deleting events. That distinction mattered to me: learn the frequency response rather than choose a fixed cutoff. Beyond Index-Only RoPE encodes event order and elapsed time through separate rotary-position dimensions or attention heads, so adjacent events need not be treated as equally close in time. Both are focused sequence-modeling changes supported here by offline, not production, results.

Semantic IDs can do more than compress

Tokens are All You Need was important to me for a reason beyond compression. Content-derived semantic IDs feed two distinct branches. One looks up embeddings trained on engagement, so items with shared codes can benefit from collaborative information learned about related items. The other reconstructs content features from the codes. I was especially interested in the first branch: content can help determine which learned parameters an item shares, while actual recommendation feedback determines what those parameters become. The reported incremental online gains are for adding reconstruction to systems that already use the collaborative branch.

UniRec takes another useful intermediate step: predict product attributes, such as categories, before generating the item code. I liked the idea of making the intended interest more explicit before choosing an item. The useful abstraction is that intermediate description of what the person might want next, rather than going directly from a history to an item identifier.

Good candidates are not enough if the next ranker does not choose them. This was the practical issue behind my interest in alignment: a stage can look better in isolation without making the final recommendations better.

Pinterest's Alignment + Accuracy derives a concrete approximation to served reward: a calibrated linear combination of overlap with the main ranker's top selections and engagement precision above a shared ranker threshold. Both terms come from logged support, avoiding full-pool rescoring. The corresponding loss trains on alignment and accuracy using their different data populations. The combined metric better predicted online winners, and the training objective improved save rate over an accuracy-only baseline.

Working through it made me see alignment as a data and measurement problem as much as a modeling one. Which ranker supplied the scores, and does the result still hold when that ranker changes? The paper's forward test is small and keeps the surrounding stages fixed. Agreement with a fixed reference ranker does not by itself establish that the improvement transfers as the downstream model evolves. The data used to define alignment is part of the problem, not just the loss function.

POEM: the ranker's belief becomes an input

The exciting part of POEM for me is that the ranker's current belief becomes part of retrieval state. It takes candidates and scores from the previous refresh and forms a partial-order sequence: retain the ordering between groups while randomizing the less reliable order within them. A second Transformer encodes this sequence alongside click history. The retriever can therefore receive an updated signal even when the person has not supplied another click.

This goes beyond asking a smaller model to imitate a teacher during training. It passes a structured summary of what the ranker currently believes into the next retrieval decision. That distinction was what made POEM exciting to me. It also raises a question about self-reinforcement: when the ranker's belief is incomplete, how does the loop avoid repeatedly favoring the same candidates? That is a question the work prompted for me, not an effect established by the paper.

Share the learning, not necessarily the serving cost

CoSID separates two jobs that need not use the same model capacity: understanding a person's history and repeatedly generating candidate item codes. A large sequence model produces a compact summary; a lightweight decoder generates semantic IDs from it, learning from both interactions and a stronger teacher's predictions. That division of expensive understanding and cheaper generation was the appeal for me. The public offline benchmarks are promising, but they leave the production memory, latency and quality trade-off to be established.

The LLM-based user persona work starts from semantically clustered watch history and generates two kinds of interests in one call: established preferences and possibilities to explore. Those teacher-generated personas train smaller Gemini Flash and Nano models; asynchronous generation and caching keep the work outside the request path. The practical contribution is both the richer profile and a way to use it without serving the large teacher.

UNIQUE uses an early-fusion backbone that produces item-code logits and ranking scores in one forward pass, training with both generation and ranking objectives. That shared computation was the appeal for me, rather than simply calling two models unified. Its reported deployment contributes candidates at retrieval; jointly training the two jobs does not mean the full serving cascade was replaced.

Pinterest's UniPinRec adds a serving connection: a sequence model produces a continuous retrieval-query embedding for an approximate nearest-neighbor index, and lightweight ranking reuses the history cache. I found it useful to separate those two ideas. Sequence-to-embedding retrieval is one capability; reusing compatible cached state across stages is another optimization on top of it.

I found the evaluation work most useful when it explained what changed in the recommendations. A better aggregate accuracy number does not reveal whether the model shifted the content mix, improved quality within each group, or traded one group's outcome for another's.

What interested me in Meta's Impression Share Prediction was the diagnostic perspective, more than the forecasting service itself. It uses score histograms, means and variances, the categories a model can score, and current delivery-capacity statistics to predict impression shares. Those inputs connect the model's predictions to the mix of content selected, which an accuracy number alone cannot explain. The scores come from the candidate's early live traffic, so this is not a fully pre-launch offline test.

Comparability from a multi-stakeholder perspective made me think more carefully about comparing outcomes across groups. It expresses a benefit relative to the gap between a disagreement baseline and an ideal reference, rather than comparing raw scores with different attainable ranges. Nash welfare then combines benefits geometrically, making proportional gains and losses matter. Those reference points are part of the modeling choice, not a neutral normalization that settles every comparison.

For me, the useful connection is with the allocation question: did quality improve within a content cohort, or did the system simply show a different mix? An aggregate gain can conceal a substantial regression for one cohort. Making those differences visible, alongside per-cohort guardrails, is more informative than treating the aggregate score as a complete verdict. The paper's comparability framework is a useful lens on that problem, not a standalone fairness guarantee.

Autonomous research still depends on evaluation

The Self-Evolving Recommendation System from Google and YouTube uses agents to propose, implement and test model changes around a shared experiment journal. A faster loop screens ideas with offline proxies; a slower loop selects what reaches live traffic. What interests me is the evaluation process around the agents. If an automated search produces many more candidates, the ability to reject misleading offline wins and learn from earlier experiments becomes even more valuable.

I encountered relatively few papers focused on evaluation, so I was glad to contribute From M Passes to One with Cheng Ding. Off-policy evaluation estimates how another recommendation policy would perform using existing logs. Our systems contribution makes the bootstrap behind its confidence intervals more efficient: one shared read feeds separately weighted accumulators for the resampling estimates. The point is to make uncertainty-aware evaluation practical at industry scale, not simply produce another point estimate.

I appreciated how seriously visitors took that systems challenge. Several senior attendees were impressed and asked detailed questions about replicating the approach. That interest in using it was the most rewarding part of my poster conversations. Better estimates tell us how a policy performs against an objective. The keynotes returned to a harder question: who chooses that objective?

Thorsten Joachims is a gift to this community. His closing keynote, Recommendation beyond Items and Engagement, connected two ideas I am especially excited about: learning the policies around individual rankings, and giving people a readable profile that can travel across recommendation systems.

Learn policies at different time scales

The first is learning at different time scales. A slower policy chooses goals or constraints, while a faster controller makes individual ranking decisions. What excited me was the possibility of learning diversity and content-balance policies from the longer-term experience, within explicit limits. The learning problem then includes the policies around individual rankings, rather than assuming every trade-off has already been specified by hand.

A profile the person can carry with them

I love the second idea: a person or their agent could carry a natural-language profile across platforms, letting different recommenders start from a shared, readable understanding of their preferences. What appeals to me is a profile richer than a list of topics, useful across different recommendation tasks rather than tied to the history of one service.

The REBEL-based training was what made this technically interesting to me. Judge a generated profile by the rankings it enables, then improve the matching embeddings alongside the profile generator. The target is not a description that merely sounds convincing. It is a description that helps another model make better recommendations.

Related papers tackle narrower parts of that problem. RECAP uses a dual-tower evaluator's matching score as the GRPO reward for training profile updates, tying the generated text to recommendation utility. The LLM-based persona work generates both established and exploratory interests from clustered history. These are concrete ways to make profiles useful, not evidence that one profile already works across platforms.

The question I am still thinking about

What happens when content is made for a person, but an agent increasingly selects, reads and summarizes it for them? I suspect the economics of attention and recommendation will change substantially. This is my reflection rather than a result from the talk: the person's preferences still matter, but the agent becomes another participant in how those preferences reach creators and platforms.

Margaret Mitchell's opening keynote raised a more personal concern for me: whether AI strengthens our judgment or gradually makes us dependent on it. I learned to understand and review code before coding agents became part of daily work. I worry about how newer developers build that depth if both writing and reviewing become approval steps. A quick approval can be a useful signal without proving that understanding or judgment improved.