GENERATIVE · 2026-10-22

Generative — Paper Daily, 2026-10-22

UGC community search for posts, questions and comments; prioritize relevance (nDCG), CTR, semantic matching, query understanding, recall, ranking and LLM applications.

4.80
LEAD STORY · Coupang / Santa Clara University

1. Evolving Semantic Indexes for Generative Retrieval

CoupangSanta Clara University

An LLM evolves executable semantic-index programs while a tiny retriever screens them. With the target training recipe fixed, NQ320K Recall@100 rises from 85.82% to 90.57%.

Key insight: Index construction and collision handling are coupled. A cheap proxy can rank index changes well, yet become misleading when the search also changes the generator objective.

📖 Read the deep dive →
4.30

2. LightCRS: Lightweight Multi-Domain Conversational Recommendation System

LightCRS caches frozen multimodal product representations offline and uses a shared query-conditioned module to compare candidates online. HR@1 improves by 35.4% over the strongest baseline.

Key insight: Separate reusable product evidence from dialogue-dependent matching. Query-conditioned fusion enables a compact shared model to serve multiple domains.

3.40

4. Self-Credit: Credit Assignment for Long-Horizon Agentic RL via Intra-group Rescaling

MetaDuke UniversityNYU

Self-Credit asks agents to cite useful earlier actions and rescales that credit across successful and failed rollouts. On BrowseComp-Plus the 8B model beats the strongest agentic RL baseline by 8.5 points.

Key insight: Self-reported credit needs cross-rollout checks. Comparing success and failure helps distinguish frequently cited actions from actions that support the answer.

3.30

5. Which Model for Which Role? Cost-Efficient Pareto Recommendation for LLM Agentic Workflows

NeurIPS 2026 WorkshopMicrosoftColumbia UniversityCornell University

CC-Gittins searches model assignments for complete agent workflows while accounting for quality and deployment cost. Across eight benchmarks it recovers a substantial part of the Pareto frontier at roughly 10% of exhaustive evaluation cost.

Key insight: Evaluate assignments end to end. Coupled evaluation and deployment costs let each trial inform both quality-seeking and cost-seeking searches.

3.15

6. DiffuPlan: Editable Planning State for Diffusion Language Model Agents

NYUNYU Shanghai

DiffuPlan maintains editable planning state and reconstructs affected steps when search evidence changes. It reaches 50.24 average F1 across four multi-hop QA benchmarks.

Key insight: Treat a plan as revisable state. Local reconstruction can replace obsolete hypotheses instead of accumulating corrections in an append-only history.

3.15

7. How Data Scales in Agentic Reinforcement Learning: Laws and Synthesis Strategies

NeurIPS 2026

The paper separates agentic training data into task, environment and trajectory counts, plus reward density, and studies scaling across models and domains using PPO and GRPO.

Key insight: Data volume hides coverage. Track the cost of new tasks, new environments and repeated trajectories separately; the recommendation connection is a proposed transfer, not an evaluated result.

3.05

9. Training the Whole Agent, Not Just the Model

Zhejiang UniversityLIGHTSPEED

HarnessAdapt-RL updates both the policy and its tools during reinforcement learning. With a 9B model it improves SWE-bench Verified pass@1 by 9.8 percentage points over GRPO.

Key insight: The deployment artifact is a model–harness pair. Temporary skills can be removed, but evolved tools and hooks remain part of the learned system.

3.05

10. DecisionPivot: A Diagnostic Benchmark for Decision Models in LLM Systems

StartLux

DecisionPivot evaluates decision correctness, probability ordering and confidence with controlled contrast groups. Two similarly ranked models score 70.1% and 44.2% contrast-group accuracy.

Key insight: Score related cases together. A model can rank probabilities well while making inconsistent decisions when a decisive detail changes.