type
Post
status
Published
date
Sep 5, 2026 05:33
slug
rec-weekly-en-2026-W36
summary
This week's recommendation systems research clusters around three main threads: generative recommendation is evolving from a single-point recall component toward industrial-grade frameworks covering ranking and reasoning; CTR modeling paradigms are reorganizing context units to align with real decision processes; and on the recall side, efficiency and cost are re-converging under the叠加 of multi-interest and multimodal approaches. Thread 1: Generative recommendation moves from "decoding items" to "unified generation and reasoning." Tencent's TGR pushes the generative paradigm into ranking, end-to-end generation, and reasoning injection — CCFormer delivers substantial gains across five A/B scenarios. Baidu's ICGR threads query-intent consistency through SID construction, SFT, and preference optimization across the full pipeline, with offline Recall@20 up 21.7%. The shared direction: generative recommendation is no longer just "replacing the index with a model" — it's starting to redraw the boundary between ranking and recall. Thread 2: Unified CTR models adopt "context" as the fundamental unit. Meituan's UniCon treats request context as a homogeneous unit, unifying the structure of history and target — online RPM up 3.09%. ByteDance's ReST demonstrates that LLM-style Transformers, after saturating on behavioral sequences, can still scale along recommendation-native design principles. Both point to the same conclusion: recommendation-specific signal noise and computational asymmetry require architecture-level redesign, not a simple transplant of NLP scaling laws. Thread 3: Cost awareness returns to the recall side. Snap's SetMIR frames multi-interest recall as set prediction, using presence scores to dynamically cut ANN queries by 33%. The same team's CAMIE replaces a fragmented I2I retrieval stack with a single multimodal encoder. Mubadala's PULSAR uses a pooled two-stage index to cut median vector retrieval latency by 15.1×. The common logic across all three: recall
tags
Recommendation Systems
Weekly
Papers
category
Rec Tech Report
icon
📚
password
priority
1
Weekly Overview
This week's recommendation systems research clusters around three main threads: generative recommendation is evolving from a single-point recall component toward industrial-grade frameworks covering ranking and reasoning; CTR modeling paradigms are reorganizing context units to align with real decision processes; and on the recall side, efficiency and cost are re-converging under the叠加 of multi-interest and multimodal approaches.
Thread 1: Generative recommendation moves from "decoding items" to "unified generation and reasoning." Tencent's TGR pushes the generative paradigm into ranking, end-to-end generation, and reasoning injection — CCFormer delivers substantial gains across five A/B scenarios. Baidu's ICGR threads query-intent consistency through SID construction, SFT, and preference optimization across the full pipeline, with offline Recall@20 up 21.7%. The shared direction: generative recommendation is no longer just "replacing the index with a model" — it's starting to redraw the boundary between ranking and recall.
Thread 2: Unified CTR models adopt "context" as the fundamental unit. Meituan's UniCon treats request context as a homogeneous unit, unifying the structure of history and target — online RPM up 3.09%. ByteDance's ReST demonstrates that LLM-style Transformers, after saturating on behavioral sequences, can still scale along recommendation-native design principles. Both point to the same conclusion: recommendation-specific signal noise and computational asymmetry require architecture-level redesign, not a simple transplant of NLP scaling laws.
Thread 3: Cost awareness returns to the recall side. Snap's SetMIR frames multi-interest recall as set prediction, using presence scores to dynamically cut ANN queries by 33%. The same team's CAMIE replaces a fragmented I2I retrieval stack with a single multimodal encoder. Mubadala's PULSAR uses a pooled two-stage index to cut median vector retrieval latency by 15.1×. The common logic across all three: recall performance is no longer measured solely by accuracy — query budget and index footprint are becoming first-class optimization targets.
Generative Recommendation and Semantic Indexing
This week's industrial progress in generative recommendation converges on one question: when semantic IDs become the sole channel for item information, how do you ensure that channel doesn't lose information critical to ranking and recall? Tencent, Baidu, and Meituan each offer answers at different levels.
TGR (Tencent) — an industrial framework covering generative ranking, end-to-end generation, and reasoning injection simultaneously. Its ranking component, CCFormer, does three things: unified feature tokenization converts user profiles, item attributes, and behavioral sequences into homogeneous tokens; a scalable Transformer backbone replaces the traditional MLP tower in fine-ranking; hierarchical sequence compression reduces computation on long behavioral sequences while preserving multi-task outputs. CCFormer delivers substantial gains across five A/B scenarios — CTR +3.57% and ad revenue +1.71% in the video scenario — and is fully deployed in two scenarios.
On the generation side, TGR converges two technical routes. BARGE and HiGR share the same starting point — encoding items as hierarchical semantic IDs, then decoding autoregressively. The difference lies in generation granularity. BARGE uses item context-aware attention and orthogonal dual-path decoding to solve the "item boundary loss" problem — the inter-level semantic drift between multi-level tokens during hierarchical SID generation. Offline Hit@5 improves 10.2%–16.9% over baseline, with CTR +0.60% after full deployment.
HiGR deserves separate attention because it answers a deeper question: does generative recommendation have to decode item by item? Its approach introduces list-level planning — first generating a plan token representing the entire recommendation list, then decoding items one by one. At inference, it uses prefix-structured SIDs and coarse-to-fine decoding. Offline slot quality improves 15.9%–21.3%, with 5× inference acceleration. The reference points here are DualGR (Kuaishou) and HiGR — both explore long/short-term interest separation and hierarchical generation in the same direction, but TGR is the first attempt to put both routes into a single production framework.
TGR's third direction is TGR-Reason — moving reasoning from the online phase to offline. It generates reason tokens for semantic IDs offline, then injects them directly during online decoding — no request-time rollout needed. Cold-start new users see Hit@1 up 477.8%, online effective consumption +1.75%, and new-user exposure-to-conversion +13.09%. The practical value here is bypassing the classic "online reasoning cost vs. reasoning benefit" tradeoff. Worth noting: TGR-Reason contrasts with PROMISE's test-time scaling logic — the latter uses a process reward model for dynamic pruning during decoding, while the former front-loads reasoning to offline. Both routes will run in parallel in industry short-term.
ICEGR (Baidu) — an intent-coherent end-to-end generative retrieval framework for e-commerce search. ICEGR's three components each address intent drift in one of the three training pipeline stages.
First, Intent-Aware SID construction. Traditional SIDs learn only from static item information; ICEGR incorporates query-intent signals into the SID learning process, so the SID encodes not just "what the item is" but also "why users search for it." Second, unified SFT with synthetic query augmentation. In e-commerce, low-exposure items lack real query-to-SID supervision; ICEGR supplements this sparse supervision with synthetic queries while unifying multiple SFT tasks under the query-to-SID objective. Third, relevance-calibrated preference optimization — injecting both query-product relevance and business signals into a margin-adaptive preference objective, preventing preference optimization from favoring high-heat items at the expense of relevance.
This "full training pipeline coverage" design directly addresses a real problem in generative retrieval deployment: if you only change the model architecture without changing training objectives, SID intent drift recurs at different stages. ICEGR decomposes this into three independently optimizable components. Offline Recall@20 up 21.7%, NDCG@20 up 26.6%; online CTR +3.52%, orders +15.96%, GMV +7.53%. The comparison point is Gryphon (industrial music service) — which addresses the misalignment between SID sequence likelihood and ranking relevance objectives. Both share the same judgment: the bottleneck in generative retrieval isn't decoding — it's correctly decomposing training objectives.
CHAP (Meituan) — approaches the same problem from a different angle. ICEGR focuses on intent consistency; CHAP focuses on semantic fidelity and inference efficiency. Its core observation: existing SIDs discretize item content, but query intent is dynamic — creating a semantic gap. And autoregressive decoding in generative recall has prohibitive latency. CHAP's Hierarchical Semantic Alignment module aligns the query latent space with the item quantization path, synchronizing semantics across multiple granularities. The personalization component fuses discrete SIDs with continuous representations — SIDs provide structural guidance, continuous representations carry fine-grained semantics.
The most engineering-valuable piece is Residual Cascading Generation: constraining multi-step Transformer decoding to a single inference pass. The approach uses residual cascading — subsequent layers only model the semantic increment that previous layers failed to express, eliminating the need for step-by-step autoregression. Validated on three public datasets, one Meituan industrial dataset, and online A/B tests: CTR +1.2%, GMV +0.8%. This "single-step generation" direction can be contrasted with xGR's serving optimization: xGR compresses generative recommendation latency through staged computation and KV cache separation, while CHAP dissolves multi-step dependencies from the model structure side. Both technical routes complete generative recall within 50ms, but start from different optimization points.
Pushing the boundaries of semantic ID fidelity: Two papers this week push SID capacity issues to finer granularity. HF-SID (Alibaba AMAP) targets LBS scenarios, noting that LLM embeddings of continuous coordinates are discontinuous — numerical differences don't reflect real geographic distance. It converts coordinates to continuous 3D Cartesian form, uses Geo-CPT and Num-CPT for type-aware encoding, and cuts average geographic distance by 95.9% with 3-token SIDs — online PV_CVR +6.74%. HypRQ-VAE moves item indexing from Euclidean to hyperbolic space — exponential volume growth naturally fits long-tail distributions, with substantial tail-item recommendation gains on Amazon Beauty, Sports, and MovieLens-1M. HF-SID extends RQ-VAE's residual quantization framework but reconstructs the encoding space; HypRQ-VAE swaps the geometric space within the same framework. Together they demonstrate: the SID information fidelity problem has shifted from "which tokenizer" to "in what space do we discretize."
Ranking and CTR Modeling Paradigms
Two industrial papers in ranking this week share a common premise: traditional CTR modeling treats behavioral sequences and non-sequential features as heterogeneous signals, but user behavior is fundamentally a homogeneous decision sequence — the only difference between history and the current request is whether the outcome has been observed. Based on this judgment, the two papers offer different unification schemes.
UniCon (Meituan) — a context-centric modeling paradigm. It treats request context as the fundamental modeling unit, organizing historical behaviors and prediction targets into homogeneous context units. Intra-context attention models local coupling between items within a unit; inter-context attention models the dynamic evolution of decision states across contexts. This is effectively a semantic segmentation along the sequence dimension — not slicing sessions by fixed time windows, but slicing units by decision context.
This design has a clear lineage to DSIN (Deep Session Interest Network). DSIN slices sessions at 30-minute intervals and uses Bi-LSTM to model interest evolution across sessions. UniCon's slicing logic is no longer time-based but decision-context-based — one request, one browse, one comparison constitutes a unit. The significance: it aligns with the real decision process in recommendation systems. In e-commerce, a user's behavior on a single page (comparing multiple items, viewing details) constitutes a complete decision context; cross-page behavioral evolution is a separate dimension. Meituan's search advertising scenario amplifies this advantage — waterfall feed contexts have higher length and item coupling than traditional scenarios. Online AUC +0.0139, RPM +3.09%, CTR +2.07%, revenue +2.95%.
ReST (ByteDance) — a recommendation-native Transformer scaling framework. Its starting point is two structural differences between recommendation and NLP. On signal quality: behavioral sequences are noisy, temporally irregular, and sparsely supervised — LLM-style Transformers hit scaling saturation faster on behavioral sequences. On computational asymmetry: a single request scores multiple candidates sharing the same user history, under strict latency budgets. ReST designs dual-gated attention, rotary temporal embedding, and training-time auxiliary objectives for the former; for the latter, it decomposes ranking into a heavy encoder plus a light decoder — user history is encoded once, and multiple candidates share prefix decoding.
This "compute-once, decode-many-times" design cuts to the fundamental tension in ranking at the system level. The paper doesn't stop at offline validation: in a one-week A/B test on an industrial advertising platform, online AUC +1.31%, core revenue metric +11.93%, completed within a 50ms P99 latency budget — and fully deployed. This thread echoes RelayGR's relay reasoning approach — RelayGR splits long-sequence generation across multiple inference stages via cross-stage relay, while ReST architecturally separates user history encoding (heavy) from candidate scoring (light). Both redraw the allocation between behavioral sequence length and latency budgets.
Taken together, UniCon and ReST represent two paths for unified CTR models. UniCon starts from the modeling unit — making "context" a first-class structural concept. ReST starts from computational asymmetry — making "shared history encoding" the first principle of system design. Both paths imply the same judgment: further CTR gains come not from stuffing more features into the same backbone, but from reorganizing input structure and computation allocation. This judgment continues the line from HSTU — which first argued that behavioral sequence modeling needs dedicated architecture design rather than directly applying LLM decoders. This week's two industrial papers provide deployment evidence under the same assessment.
Recall and Vector Retrieval Optimization
Industrial progress on the recall side manifests as subtraction rather than addition: fewer redundant queries, fewer encoders, smaller index footprints. Three efforts compress recall costs across multiple dimensions.
SetMIR (Snap) — multi-interest recall shifts from "fixed-number embeddings" to "dynamic-number set prediction." Traditional multi-interest recall (e.g., MIND's multi-vector approach) has two structural problems: interest collapse — multiple embeddings learn the same interest; and static dispatch — even when certain interest vectors aren't needed, serving still issues the same number of ANN queries. SetMIR uses Hungarian matching for one-to-one query-to-interest assignment, forcing each query to learn a distinct interest. A presence head learns which queries are active for the current request. At serving time, presence scores plus query-level NMS only issue ANN queries for active, non-redundant interests.
On numbers: compared to a 4-learned-interest multi-interest recall baseline, SetMIR wins on every metric while cutting ANN queries per request by 33%. Deployed as a new recall source in Snap's DPA production stack, overall CVR +3.1%. Compared to an item-to-item recall source with the same embeddings and same index, CTR up 44% and CVR up 51%. This comparison is especially telling — identical item embeddings, identical ANN infrastructure, but changing query construction from "given an item, find similar" to "given user history, decode an interest set" produces an order-of-magnitude gap in conversion outcomes.
CAMIE (Snap) — co-participation-aware multimodal item embeddings. DPA's I2I recall previously had two problems: visual, text, and multimodal encoders each did their own job, fragmenting the retrieval stack; and pure content training meant embeddings never aligned with the co-participation signals that actually drive conversion behavior. CAMIE's solution: unify the retrieval stack into a single multimodal encoder, fine-tuned on co-participated item pairs. The most noteworthy methodological detail: it reuses the LLM/MLLM native multimodal interface without modifying the backbone for recommendation tasks — image visual tokens and item metadata text tokens go directly into the same backbone.
On results, CAMIE exceeds the strongest commercial multimodal embedding model on Recall@10, and the same checkpoint supports text-only retrieval with minimal quality loss. After online replacement of two content-based I2I encoders: CTR +0.390% and CVR +10.832% vs. the multimodal control; CTR +18.958% and CVR +13.12% vs. the text control. Compared to general image-text models like CLIP and SigLIP, CAMIE's key difference is the training signal — not "are image and text paired" but "are two items co-participated in the same user journey." This signal alignment logic complements HyMiRec's decoupling of coarse-grained interests from lightweight recommenders and fine-grained interests from LLMs — both answer "at what granularity should behavioral signals be injected into content representations."
PULSAR (Mubadala Investment Company) — a pooled two-stage late-interaction index. It targets high-density visual document retrieval in financial research — pitch decks, board packs, and other chart-dense, frequently revised materials. Traditional OCR-plus-chart-annotation pipelines have high update costs and lose chart details. PULSAR uses a frozen ColPali-style backbone to index page images directly. The core engineering innovation is a pooled two-stage late-interaction index: compact page summaries for initial screening, then precise MaxSim re-scoring on finer pooled representations in the refinement stage.
The performance numbers deserve close reading. On ViDoRe V3, compared to the unpooled configuration: median vector retrieval latency down 15.1×, with NDCG@10 and Recall@10 losses both under 0.01 absolute points. Production median latency is 156ms. Under concurrent load, the pooled index achieves roughly 88× the QPS of unpooled. The engineering value: it makes an explicit priority ordering of late-interaction's two-stage cost distribution — first-round retrieval must be cheap, final re-ranking must be accurate. This complements Col-Bandit's query-time pruning direction: Col-Bandit reduces MaxSim computation at runtime; PULSAR pools expensive representations at index time.
Two phenomena worth noting: First, recall-side gains are shifting from pure algorithms to joint algorithm-engineering optimization — SetMIR's 33% query reduction and PULSAR's 88× QPS both come from joint design, not single-model improvements. Second, the unification trend in multimodal recall is clear — CAMIE replaces three encoders with one, aligning with BiListing's approach at Airbnb of replacing a multimodal stack with a single embedding vector.
LLM and Agent-Enhanced Recommendation Reasoning
No large-scale deployed systems emerged in the LLM/Agent direction this week. Several papers focus on the boundary problems of inference efficiency, simulators, and distillation. Compared to previous weeks' "LLM directly does recommendation" narrative, this week's papers mostly answer a reverse question: under what conditions and in what form can LLM reasoning capabilities enter recommendation systems affordably?
Inference cost is the core constraint. SelfDR confronts it directly — distilling reasoning-augmented predictions back into a direct recommendation model. A teacher model first trains a reasoner with downstream performance as reward, generating rationales added to the input. A same-backbone student model learns via self-distillation, with a dynamic weighting strategy adjusting distillation strength during training. The advantage: no external model dependency — teacher and student share the same backbone. Compared to ReRec's approach of fine-tuning LLMs with reinforcement learning to generate reasoning, SelfDR's difference is treating reasoning as "training-time scaffolding" rather than "a required inference-time step." This is a direct challenge to the implicit assumption that reasoning must happen online.
Another direction: replacing real online feedback with simulators. WMG-RL (ByteDance with academic partners) proposes a User Engagement World Model (UEWM) — treating recommended items as agent actions, user feedback as environment observations, inferring user-specific dynamics from historical engagement data, then applying them to candidate items. The downstream policy proposes multiple candidate items for the same history; UEWM predicts engagement feedback in parallel, converted to dense rewards for policy optimization. Experiments show a 1.7B student policy can match or exceed much larger LLMs. This traces back to language world model thinking, but applying it to user engagement simulation still needs to address simulator bias — if UEWM's predicted engagement systematically diverges from real feedback, policy optimization amplifies that divergence.
In conversational recommendation, one methodological reflection is worth noting this week. CLEAR identifies a systematic answerability gap in conversational retrieval — topically relevant passages aren't necessarily passages supporting the correct answer. CLEAR uses entailment distillation to transfer answer-passage entailment supervision to a cross-encoder re-ranker, enabling the re-ranker to distinguish "answer-supporting passages" from "topical distractors" without seeing the answer at inference. A companion passage-based abductive recall module uses LLMs to reverse-engineer answerable queries from passages, pulling low-similarity-but-answerable passages into the candidate pool. On TopiOCQA, QReCC, and out-of-domain TREC CAsT, CLEAR consistently improves top-ranked precision — with larger gains on noisier conversational topics. This "topical relevance ≠ answerability" distinction maps directly onto recommendation's old problem of "relevance ≠ click."
LLM4AIGQ (Alibaba) — using LLMs for e-commerce guided query (AIGQ) generation. On a nearline-generation, online-read architecture, it first segments interests by user profile and historical behavior, then infers consumption intent for each sub-interest, and finally generates corresponding guided queries. The training pipeline is SFT + RL + DPO post-training with multi-level reward design. Deployment uses nearline generation to satisfy latency constraints. Validated in both offline evaluation and online A/B tests on Alibaba e-commerce search. The multi-level reward design continues RecThinker's multi-step tool-call reward decomposition, but moves the target from recommendation itself to query generation.
Taken together, these efforts show LLM recommendation research shifting from "can LLMs improve results" to "how to control the marginal cost of bringing LLMs into recommendation systems." SelfDR moves reasoning out of the online path; WMG-RL moves feedback into a simulator; LLM4AIGQ moves generation into nearline. Three "moves" pointing at one goal — getting LLM semantic understanding into the recommendation pipeline at controllable cost.
Directions to Watch
Long-tail and cold-start are becoming generative recommendation's main battleground. Multiple efforts this week treat long-tail performance as a core metric: TGR-Reason lifts cold-start new-user Hit@1 by 477.8%, ICEGR uses synthetic queries to supplement supervision for low-exposure items, HypRQ-VAE improves tail-item representation fidelity in hyperbolic space, and MERIT handles long-tail matching across 250k+ interest categories. The concentration is structural — SID encoding quality directly determines model perception of low-frequency items, and long-tail items are precisely where SIDs distort most. Worth tracking: now that TGR has proven reason tokens can be injected offline, will long-tail cold-start become generative recommendation's first stable industrial value anchor?
Index efficiency is becoming the core metric for retrieval optimization. SetMIR's query-level NMS, CHAP's residual cascading single-step decoding, PULSAR's pooled index, HF-SID's 3-token SID — each compresses one link in the generation or retrieval chain. But one question remains unanswered: if these techniques were combined in a single system, would gains stack or interfere? CHAP's single-step generation paired with PULSAR's pooled index could simultaneously reduce decoding steps and re-ranking costs — but how their error propagation would interact has no systematic research yet.
Behavioral sequence "context segmentation" is moving from engineering heuristics to theoretical frameworks. DSIN's time-based session slicing is an engineering heuristic. UniCon's decision-context-based unit slicing begins to have theoretical consistency. ReST's shared-prefix decoding extends slicing logic to computation allocation. The evolution suggests: the next increment in behavioral sequence modeling may not come from better attention mechanisms, but from more accurate modeling of sequence structure — what constitutes a complete decision unit, and how units evolve. This cross-validates with Modern Transformers' finding of "head-level functional differentiation" in Head-wise Hybrid architectures: when internal model organization aligns with task structure, efficiency and effectiveness improve together.
Paper Roundup
Generative Recommendation and Semantic Indexing
TGR — Tencent proposes an industrial framework unifying generation and reasoning; CCFormer video CTR +3.57%, ad revenue +1.71%; BARGE offline Hit@5 up 10.2%–16.9%; HiGR 5× inference acceleration; TGR-Reason cold-start new-user Hit@1 up 477.8%.
ICEGR — Baidu proposes an intent-coherent generative retrieval framework; offline Recall@20 +21.7%, online CTR +3.52%, GMV +7.53%.
CHAP — Meituan proposes cross-component hierarchical semantic alignment with residual cascading generation for single-step inference; online CTR +1.2%, GMV +0.8%.
HF-SID — Alibaba AMAP uses Geo-CPT/Num-CPT to restore geographic and numerical fidelity; average geographic distance down 95.9%, online PV_CVR +6.74%.
HypRQ-VAE — Academic team proposes the first hyperbolic-space item indexing framework, using exponential volume expansion to fit long-tail distributions; substantial tail-item recommendation gains.
EPIC — Academic team introduces explicit item-level posterior conditioning in SID diffusion recommendation, freezing pretrained backbone with no extra decoding passes; consistent gains across four Amazon benchmarks.
TAAL — Academic team proposes temporal autoregressive alignment to mitigate early beam pruning; NDCG@10 up 6.7%–39.5%.
MERIT — Amazon proposes permutation-invariant multi-objective loss to mitigate generative XMC exposure bias; global recall +11.9%, production A/B user conversion +0.26%.
CoGR — Apple has an LLM directly generate query-side and item-side keywords with co-evolutionary RL optimizing both ends; internal dataset F1 +10.9%, WANDS +36.1%.
Joint Embedding-Codebook — Alibaba jointly trains embeddings and codebooks with same-product-cluster supervision to mitigate two-stage cascading errors.
Ranking and CTR Modeling Paradigms
UniCon — Meituan proposes context-centric unified modeling; online AUC +0.0139, RPM +3.09%, CTR +2.07%, revenue +2.95%.
ReST — ByteDance proposes a recommendation-native Transformer scaling framework; online AUC +1.31%, core revenue +11.93%, fully deployed.
HubMixer — Kuaishou proposes latent hub feature interaction architecture; resume submission conversion +5.48%, fully deployed.
PRIME — Ant Group/Alibaba proposes dense-anchored residual MoE to mitigate subgroup optimization competition; median AUC gain +0.0022/+0.0066.
SwapRec — Albatross AI replaces cold-item interactions with most-similar warm items during training; substantial cold-item accuracy gains across three recommendation domains.
TS-SSM — Academic team proposes a two-sided state space model for non-random review feedback; Recall@20 relative gain 14.8%–18.8%.
Adaptive Doubly Robust — Academic team proposes an OPE estimator for ranking combining adaptive importance weighting with reward regression; MSE consistently outperforms AIPS in synthetic experiments.
Bandits in Prod — Tiime formalizes inference-time configuration optimization as an infinite-armed bandit, proposes IMABO/IMOSS with proven cumulative quantile-regret bounds.
Recall and Vector Retrieval Optimization
SetMIR — Snap models multi-interest recall as set prediction; presence scores dynamically cut 33% of ANN queries; online overall CVR +3.1%.
CAMIE — Snap fine-tunes a multimodal encoder with co-participation signals, unifying the I2I retrieval stack; online CTR +18.958%, CVR +13.12%.
PULSAR — Mubadala deploys a pooled two-stage late-interaction index; median retrieval latency down 15.1×, production QPS up ~88×.
PAO — Alibaba proposes positive-only advantage gradient updates to avoid embedding geometric collapse under frozen-index constraints; substantially outperforms standard RL and distillation baselines.
MIDR — Bloomberg proposes an index-time multimodal reasoning framework replacing serving-time visual late-interaction; average nDCG 0.6219, index memory reduced ~9×.
Closed Forms and Synthetic Twins — Academic team proposes a theoretical framework predicting ANN recall from embedding statistics; prediction error under 0.03 on million-document corpora.
Surface-Form Bias — Academic team reveals surface-form bias in embedding retrieval on structured tasks; math-domain Hit@1 as low as 0.0%.
RePair — Academic team converts retrieval failures into counterfactual hard samples with LLM-guided edits for hard negative pairs; data efficiency up 26%–75%.
Edge Spectrum — Academic team proves strong and weak edges in choice-derived item graphs encode different relations, formalizing sign mismatch between smoothing operators and ranking gradients.
MULTI3IR — Academic team builds a multi-view multi-domain multimodal retrieval benchmark (104.9K queries), proposing SPIN for view coverage improvement.
ITER — Academic team proposes an interaction-aware retriever trained on agent trajectory signals; InfoSeek-Eval +7.5%, BrowseComp-Plus +13.5%.
AMUR — Academic team proposes information-theoretic selective modality-interest alignment; Recall@20 up ~5%–8% across three real datasets.
LLM/Agent-Enhanced Recommendation Reasoning
LLM4AIGQ — Alibaba uses LLMs for multi-interest mining and guided query generation with SFT+RL+DPO post-training; validated in offline evaluation and online A/B tests.
SelfDR — Academic team proposes a self-distillation framework transferring reasoning-augmented predictions to a direct recommendation model with zero inference-time overhead.
WMG-RL — ByteDance with academic partners proposes a user engagement world model generating counterfactual rewards; 1.7B policy matches or exceeds much larger LLMs.
CLEAR — Academic team uses entailment distillation to shift conversational retrieval from topical relevance to answerability; largest gains on topically noisy conversations.
DREAMS — Academic team uses dual-node MCTS tree structure to model preference evolution and exploitation in conversational recommendation.
AgentMMRec — Academic team uses LLMs as knowledge integrators and utilizers, converting generated knowledge to graph structure before recommendation; consistent gains across three Amazon datasets.
Beyond Ranking Accuracy — Walmart Global Tech evaluates the ranking value of LLM repurchase explanations; LLMs shouldn't serve as standalone rankers but can serve as validated explanation components.
Other
Modern Transformers — Academic team proposes RFIS/RPD metrics revealing a complete taxonomy of retrieval and positional heads in RoPE Transformers, designing the HwH hybrid architecture accordingly.
RATL — Academic team proposes residual retrieval and feedback correction, shifting retrieval targets from target values to residuals of base model historical predictions.
SOMTab — Academic team proposes a Set-Order Mamba architecture for tabular in-context learning, separating representation construction from query-conditioned retrieval for a better efficiency-accuracy tradeoff.
PMFRec — Academic team proposes personalized multi-view federated cold-start recommendation with orthogonal objectives eliminating cross-view redundancy and substantially reducing communication overhead.