Paper Notes, May 2026: Five Papers on Ranking, Training, and Retrieval

AI-assisted reading notes, drafted with an AI assistant and not yet verified line by line against the original papers. Corrections are welcome.

Compiled by: Chris Yan | Date: 2026-06-05 | Papers: 5

Paper Index#

Paper 1: CrossDistil — Cross-Task Knowledge Distillation#

Core Problem#

In multi-task learning (MTL), the predictions of different tasks may carry complementary ranking information, but directly using one task’s soft labels to teach another leads to task conflict (different tasks may rank the same item in contradictory ways).

Method#

graph TB
    subgraph CrossDistil["CrossDistil Framework"]
        AUG["1. Task Augmentation<br/>Introduce auxiliary ranking tasks<br/>Learn fine-grained ranking<br/>relations via quadruplet loss"]
        CAL["2. Calibrated KD<br/>Platt Scaling calibration<br/>Align prediction magnitudes<br/>across tasks"]
        ERR["3. Error Correction<br/>Clamp incorrect teacher predictions<br/>Speed up convergence"]
    end

    AUG --> CAL --> ERR

Three key contributions:

1. Augmented Ranking Tasks — Quadruplets (pos_A, neg_A, pos_B, neg_B) are constructed into auxiliary ranking tasks, avoiding the ranking conflicts caused by direct cross-task distillation
2. Calibrated Distillation — Platt Scaling (r̃ = P·r + Q) aligns the teacher’s logit magnitudes before KD, removing the bias caused by differing positive rates across tasks
3. Error Correction — When a teacher prediction contradicts the hard label, the logit is clamped, so an inaccurate early-stage teacher does not mislead the student

Key Findings#

• On the TikTok and WeChat datasets, CrossDistil achieves the best AUC and Multi-AUC over all MTL baselines (Shared-Bottom, Cross-Stitch, MMoE, PLE)
• The auxiliary ranking tasks contribute more than calibration and error correction
• Direct cross-task KD (without the augmented tasks) actually degrades performance

Implications for Practice#

In industrial two-tower retrieval models, adding an auxiliary alignment loss between the sequence encoder and the item embeddings is a common way to add supervision, and it is essentially a form of cross-module knowledge transfer. CrossDistil’s calibration idea suggests checking that the two embedding scales match before applying a cosine penalty; otherwise, direct alignment may be biased.

Paper 2: Disaggregated Multi-Tower (DMT) — Topology-Aware Training#

Core Problem#

In recommendation model training, embedding AlltoAll communication accounts for 27.5% of training time (Figure 1). The root cause is a mismatch between the model architecture (flat, global interactions) and the datacenter topology (hierarchical, with intra-host NVLink » inter-host RDMA).

Method#

graph TB
    subgraph DMT["Disaggregated Multi-Tower"]
        SPTT["1. SPTT (Semantic-Preserving Tower Transform)<br/>Decompose global embedding placement<br/>into local towers<br/>Exploit intra-host NVLink"]
        TM["2. Tower Module<br/>Hierarchical feature interaction<br/>within each tower<br/>Compress cross-host traffic"]
        TP["3. Tower Partitioner<br/>Learn feature grouping<br/>Place semantically similar features in the same tower<br/>+ load balancing"]
    end

    SPTT --> TM --> TP

Core idea: Split a single flat model into multiple towers, each corresponding to a group of semantically related features and placed within a single host. Communication within a tower uses high-bandwidth NVLink; cross-host AlltoAll between towers happens only when necessary.

Key Numbers#

Implications for Practice#

For recommendation models with a very large number of sparse features, DMT’s SPTT + Tower Module approach applies directly when scaling to more GPUs: group sparse features by semantics, partition them automatically with the Tower Partitioner, and reduce training communication overhead.

Paper 3: Enhanced Twin-Tower — Early Interaction + Ranking Loss#

Core Problem#

Conventional twin-tower models have two limitations: (1) query and document are encoded fully independently, with no early interaction; (2) training uses a loss on absolute similarity scores, which tends to overfit to a specific distance threshold.

Method#

graph TB
    subgraph Model["Enhanced Twin-Tower"]
        EARLY["Early Interaction Layer<br/>Inject a query pseudo-representation<br/>during document encoding<br/>(cross-attention)"]
        RANK["Ranking-Optimized Loss<br/>temperature-scaled<br/>contrastive loss<br/>Optimize relative order, not absolute distance"]
    end

    EARLY --> RANK

Two improvements:

1. Early Interaction — Cross-attention is added to the document encoder to generate a pseudo-query representation from document content, so the document embedding already contains query information at encoding time. MRR improves by 9.2%
2. Ranking-Optimized Loss — A temperature-scaled contrastive loss focuses on relative ranking rather than absolute similarity scores and is more robust to label noise. F1 improves by 14.6%

Key Findings#

• On three QA datasets (NQ/TQA/WQ), Top-20 accuracy improves by 20.3% over BM25
• Inference latency is only 17 ms (retrieving over 1K candidates), on par with DPR
• The two components are synergistic: using them together works better than either alone

Implications for Practice#

The idea is similar to cross-attention inside user behavior sequence encoders in recommendation models: both inject query-side information at encoding time. This paper, however, targets NLP retrieval rather than recommendation. In recommendation, an auxiliary loss that directly optimizes embedding alignment (rather than only softmax CE) can likewise be viewed as a ranking-aware loss.

Paper 4: DHEN for CVR — Pinterest’s Conversion-Rate Prediction in Practice#

Core Problem#

CVR prediction faces three challenges: (1) labels are extremely sparse (orders of magnitude fewer than CTR labels); (2) labels are delayed (off-site conversions arrive late); (3) labels are noisy (attribution is imprecise).

Method#

graph TB
    subgraph DHEN_CVR["DHEN for CVR"]
        MTL["Multi-Task Learning<br/>Joint training of CTR + multiple CVR tasks<br/>CTR head provides abundant training signal"]
        
        subgraph Architecture["2-Layer DHEN"]
            L1["Layer 1: MLP + Transformer"]
            L2["Layer 2: MLP + MaskNet"]
            L1 --> L2
        end
        
        SEQ["Sequence Modeling<br/>5 types of behavior sequences<br/>(search, organic, ads,<br/>matched conv, attributed conv)"]
        SSL["Self-Supervised Loss<br/>Predict future actions in the sequence<br/>(Next Action Loss)<br/>Mitigate label sparsity"]
        PARETO["ParetoNet Search<br/>Performance-throughput<br/>Pareto-optimal search"]
    end

Key design decisions:

Online Results#

Key observation: CTR drops while CVR rises, indicating the model more precisely identifies users who will actually convert instead of merely chasing clicks.

Implications for Practice#

Pinterest’s DHEN architecture is an ensemble of multiple feature-crossing modules. Its self-supervised next action loss addresses the problem of giving the sequence encoder additional supervision. A common alternative is cosine alignment, whereas they use InfoNCE for next-item prediction. They also find that sequence features matter more than pre-trained embeddings, which suggests that investing in the sequence encoder is the right direction.

Paper 5: UniNote — Unified Embedding + RL Reranking#

Core Problem#

Industrial I2I (item-to-item) retrieval faces three problems: (1) a granularity mismatch between global representations and local retrieval; (2) a disconnect between the embedding and ranking pipelines (retrieve-then-rerank is inefficient); (3) heterogeneous fusion of multimodal content (text + images + video + OCR).

Method#

graph TB
    subgraph UniNote["UniNote: Two-Stage Training"]
        subgraph Stage1["Stage 1: Contrastive SFT"]
            DATA["10 I2I retrieval tasks<br/>+ multi-granularity<br/>hard negative mining"]
            MLLM["Qwen3VL-8B → embedding<br/>last token as representation<br/>+ MRL (64/512/1024/4096)"]
            LOSS1["JS Divergence<br/>(embedding dist vs MLLM soft label)"]
        end

        subgraph Stage2["Stage 2: RL Reranking (GRPO)"]
            REWARD["Hierarchical Reward:<br/>1. Irrelevant Note Penalty<br/>2. Base Relevant Reward<br/>3. Absolute Position Reward<br/>4. Relative Order Reward"]
            GRPO["GRPO optimization<br/>directly on embedding space"]
        end
    end

    Stage1 --> Stage2

Key contributions:

1. Unified retrieval-ranking framework — The conventional setup uses two models: an embedding model (retrieval) plus a reranker (ranking). UniNote uses a single model for both retrieval and ranking; the RL stage gives the embedding space ranking ability directly
2. Contrastive SFT — Turns an MLLM into an embedding model (using the last token) and uses JS divergence to align the embedding distance distribution with the MLLM soft-label distribution. Combined with MRL to support multiple deployment dimensions
3. GRPO Reranking — RL training with a 4-level hierarchical reward directly optimizes ranking quality in the embedding space. No separate reranker model is needed
4. Hard Negative Mining — A three-tier strategy: similarity-based (global candidate pool) + rule-based (different elements within the same note serve as negatives for each other) + counterfactual (replacing part of the content)

Key Results#

Implications for Practice#

This paper is highly relevant to VLM embedding work. A few directly transferable ideas:

1. MRL — Use a Matryoshka loss to support deployment at multiple dimensions
2. GRPO for ranking — Use RL to directly optimize ranking quality in the embedding space, without a separate reranker. Retrieval models trained only with contrastive learning could consider adding an RL stage
3. Three-tier hard negative mining — Instead of relying only on in-batch negatives, try similarity-based + rule-based + counterfactual negatives
4. JS divergence vs cosine penalty — UniNote uses JS divergence to align two distributions (embedding vs MLLM soft label), which is more principled than a plain cosine penalty

Cross-Paper Comparison#

Theme: Adding Extra Supervision to Embedding / Sequence Models#

Theme: Efficiency Optimization for Large-Scale Recommendation Models#

Takeaways for Industrial Recommender Systems#