GENERATIVE · 2026-10-20
Generative — Paper Daily, 2026-10-20
Community search over short user-generated posts, questions, answers, and comments. Main metrics: nDCG and CTR. Priorities include semantic matching, query understanding, retrieval, ranking, and LLM applications in search.
3.65
Some applications need candidate probabilities rather than generated explanations. Chinese Decision normalizes candidate-letter logits from a single forward pass for Boolean, categorical, and ordinal decisions. Its completed 0.8B model reaches 72.24% general-domain accuracy on CJ-Bench.
Key insight: Expose candidate probabilities directly when a task only needs a finite decision. This avoids autoregressive answer generation, while calibration and individual prediction changes still require their own evaluation.
3.60
MetaA single agent mixed up rules when onboarding data columns across recommendation pipelines. Meta moved routing and reliability checks into a thin orchestrator and isolated domain reasoning in specialist agents. Configuration preparation fell from three to four days manually to under an hour.
Key insight: Keep workflow sequencing deterministic and load domain knowledge only in the relevant specialist. Recombining existing workflows can add capabilities without expanding the central agent’s context.
3.55
QQilu University of Technology
UTS Separate shared experience from local prescribing patterns before generating a medication set. HEM-Med retrieves both memory levels and enforces safety and formulary constraints during decoding. It reports an F1 score of 0.4532 on the multicenter eICU dataset.
Key insight: Keep transferable evidence separate from local preferences. Applying both compatibility and availability constraints during set generation is a useful design for constrained recommendation; the evaluation here is clinical, rather than consumer recommendation.
3.45
Generative-search optimization studies often allow different source edits, making their results difficult to compare. This position paper organizes edits by layer, action, and target signal, then specifies controlled comparisons. It proposes a protocol rather than reporting a benchmark performance gain.
Key insight: Control edit permissions and optimization budgets before comparing citation gains. A source can become more visible while degrading the original task, so visibility and task quality need separate measurements.
3.45
IIndependent Researcher
Text revision often regenerates large amounts of unchanged context. Causal Edit Serialization emits pointer-addressed edit programs inside a causal token stream while retaining the original context. The localization components add approximately 0.34–0.41% parameters at the tested depths.
Key insight: Represent localized revision as an executable program. Separating edit content from source addresses makes transformations inspectable and lets the system retain source context and cached states.
3.00
Masked diffusion commits tokens too early to revise poor choices. LED keeps an editable proposal state and refines future predictions before committing them. Its 7B model improves the eight-benchmark average by 11.4 points over standard masked-diffusion supervised fine-tuning.
Key insight: Separate proposal updates from final token commitment. This preserves iterative revision while retaining mask-based training, although extra proposal updates also consume inference compute.
3.00
Diffusion decoding must predict both tokens and the order in which positions are committed. SPOT learns that order, proposes a tree of continuations, and verifies branches in parallel. Reported peak throughput is 3.2 times standard decoding.
Key insight: Learn the commitment order as well as token content. Prior output distributions contain useful draft tokens, but a verifier still needs to check which proposed state is actually the next target state.
3.00
Observed context can be wrong even when a diffusion model treats it as fixed. Region-dependent corruption teaches the model to detect and revise suspicious context. With 20% uniform-random corruption, infilling token F1 improves by 6.4 points.
Key insight: Observed tokens need not be fully trusted. A small revision allowance helps with obvious corruption, but gains on plausible errors are weaker and were not statistically significant in the reported setting.
3.00
Sparse attention can spend much of its savings recomputing token indices across layers. Prism Attention shares coarse proposals and refines them locally for each layer. It reports 3.4-times end-to-end decoding speedup over FlashAttention at a 128K context length.
Key insight: Share candidate regions across layers while allowing local refinement. Stable high-attention cores can be reused, with computation concentrated on changing boundaries instead of recomputing all indices.
3.00
Jane Street
NVIDIAMamba-2 can lose time moving intermediate state between kernels. The authors use cuTile to fuse tiled computation and optimize memory layouts for the forward pass. On B300, the kernels run 1.14–1.82 times faster than the Triton reference.
Key insight: Separate inter-chunk state propagation from intra-chunk computation and keep temporary matrices in registers. These are forward-kernel results; backward computation and other pipeline costs determine end-to-end gains.
NeurIPS 2026