GENERATIVE · 2026-10-16
Generative — Paper Daily, 2026-10-16
Community search over short user-generated posts, questions and comments. Priorities include semantic matching, query understanding, retrieval, ranking and LLM applications; key metrics are nDCG and CTR.
3.95
Changing who employs an assistant changes how it treats sponsored listings. Controlled hotel-choice experiments hold prices and reviews constant while varying delegation and sponsorship. In one condition, the sponsored-choice share differs by 22.2 percentage points between roles.
Key insight: Delegation is a ranking variable. Hold candidates fixed and vary the principal and sponsorship label when auditing a conversational recommender; disclosure alone does not determine how the model treats paid placement.
3.50
SemICD replaces raw diagnosis codes with hierarchical semantic identifiers. It clusters code descriptions and trains bidirectional description-to-ID tasks. Across four LLMs and two medical datasets, Micro-F1 gains range from 1.34 to 8.00 points.
Key insight: A useful identifier prefix should reflect task semantics. Bidirectional description/ID training gives discrete tokens meaning; the evidence here concerns medical coding, with identifier design offering the transferable idea.
3.35
ReaDy reads a hybrid linear-attention model's recurrent state to provide position-specific context for a speculative drafter. Low-rank reads and fused operators limit overhead. On one B200, it achieves 3.16× mean decoding speedup over autoregressive Qwen3.8-27B.
Key insight: Committed recurrent states offer context beyond the last hidden feature. For hybrid linear-attention serving, compare state-aware drafting against a hidden-feature-only drafter and measure whether longer acceptance offsets the extra memory traffic.
3.35
TreeSpark trains a drafter on the candidate trees its verifier will actually inspect. It replays expansion, pruning and budgeted packing using existing target distributions. Mean single-request throughput improves 15.1–25.0% over DSpark across evaluated targets and suites.
Key insight: Train on the tree that survives inference, rather than only on a reference chain. Existing tree verifiers can reuse their branching budget and pruning policy during drafter training so supervision follows retained alternatives.
3.35
SpecAgent lets a small model draft grounded tool trajectories and a strong model verify whole segments. Read-only calls execute during drafting; mutations wait for review. On τ²-Bench, wall time is roughly halved while success stays within one point of the strong model.
Key insight: Grounded trajectory review replaces repeated strong-model generation with parallel verification. A tool-using recommendation assistant would need to preserve real observations and place a verification boundary before external mutations.
3.35
AutoSFT gives agents a broader action space for training-data research: reweight, prune, repair, replace and synthesize. Execution, analysis and synthesis have separate roles. Under fixed training conditions, analysis components raise the best IFEval score by 6.5 points.
Key insight: Data research needs more options than mixture weights. For a recommendation assistant, rare-intent synthesis, label repair and duplicate removal can be separate actions, compared through retraining from the same base model each round.
3.35
JevAgentHarness delegates selected choices to a cheaper decision model and falls back to the LLM below calibrated thresholds. Training includes states reached after agent mistakes. On tasks similar to training, evaluated 4B models reduce agent cost by 81.1–89.0%.
Key insight: Isolated decision accuracy does not predict end-to-end task success. Train on downstream states caused by mistakes and calibrate delegation thresholds on actual trajectories, where an inexpensive wrong choice can create costly rework.
3.25
One Block Ahead denoises the next text block alongside the current one. It promotes tokens together with their sampler state, including denoising time and self-conditioning. On DiffusionGemma, mean single-request denoising throughput rises 16.3% across ten benchmarks, with a small accuracy regression.
Key insight: Parallel block generation must move sampler state along with tokens. The gain depends on the cost of a paired-block forward pass; the paper's break-even criterion connects local parallelism to actual latency.
3.25
This work streams the tokenization pipeline instead of waiting for each stage to finish a complete input. It formalizes output delay and supplies theoretical bounds plus practical implementations. Time to first token improves 7.9–34.8× over tiktoken in reported tests.
Key insight: First-token delay and total throughput need separate budgets. Streaming query processing and long-document ingestion should measure both; approximately constant preprocessing memory also limits spikes from large inputs.
3.10
ixi-sLM combines vocabulary trimming, a call-adapted tokenizer and task-aware mixed-precision quantization for mobile call assistance. One model powers call summarization and answering. The deployed model occupies 803 MiB and serves several million subscribers.
Key insight: Vocabulary and output layers deserve their own memory budget. Frequency-based trimming and real-task quantization calibration can shrink both model storage and input length; this production report demonstrates the approach on mobile NPUs.
NeurIPS 2026 Workshop
LG Uplus