# paper-forge daily report — 2026-09-28

19 new papers screened; 10 analyzed in depth; 0 reproduction runs.

## Reproductions

- none today (reimplementation is opt-in: `forge daily --reimplement`)

## Top candidates

### [Bootstrapping Conversational Recommendation Agents At Spotify: Synthetic Data Generation and Self-Improvement Loops](https://arxiv.org/abs/2609.30297)
`2609.30297` · score 45.8 · analysis: heuristic · feasibility 6/10

Conversational recommendation agents are a new paradigm for content discovery, enabling users to express complex intents through natural language (e.g., "recommend Italian indie artists I haven't heard before"). A central challenge in building such agents is optimizing agent planning -- deciding how to select, sequence, and invoke tools -- particularly in cold-start settings where real user interactions are not yet available.

Rubric evidence: + toy/synthetic experiments ("Synthetic"); + small/synthetic data ("Synthetic"); + claims simplicity ("simple"); - system/framework (many moving parts) ("agents are a new paradigm for content discovery, enabling users to express complex intents through natural language (e.g., "recommend Italian indie artists I haven't heard before"). A central challenge in building such agents is optimizing agent planning -- deciding how to select, sequence, and invoke tools -- particularly in cold-start settings where real user interactions are not yet available. We introduce a pipeline for multi-turn synthetic data generation and a self-improvement loop to address this challenge. The synthetic data pipeline transforms single-turn prompts into realistic multi-turn conversations, enabling systematic evaluation before launch. The self-improvement loop combines variance-based contrastive optimization with iterative refinement through a coding agent, automatically identifying and fixing planning and tool"); no public code found

### [RecToolBench: Benchmarking Recommendation-Specific Tool Orchestration under Fuzzy User Intent](https://arxiv.org/abs/2609.30717)
`2609.30717` · score 42.1 · analysis: heuristic · feasibility 4/10

Recent advances in agentic recommender systems are shifting recommender systems from passive filtering engines to instruction-following agents that use external tools to resolve user intent. However, existing benchmarks often assume explicit user intent, simplified tool environments, or isolated function calls, leaving realistic tool orchestration for recommendation underexplored.

Rubric evidence: - LLM-based ("LLM"); + public benchmarks ("benchmarks"); + explicit algorithm/pseudocode ("Algorithm 1"); - system/framework (many moving parts) ("agentic"); + code: https://github.com/ShawnChenn/RecToolBench

### [CG-Probes: Recovering Guardrail Directions from Patient Query Embeddings](https://arxiv.org/abs/2609.31062)
`2609.31062` · score 41.6 · analysis: heuristic · feasibility 6/10

Patient-facing AI assistants promise valuable support to patients, but incoming queries can pose medical risks. To create guardrails, we work with oncologists to define three ordinal risk axes: Medical Urgency, Psychological Urgency, and Topic Sensitivity.

Rubric evidence: - LLM-based ("LLMs"); + toy/synthetic experiments ("synthetic"); + small/synthetic data ("synthetic"); + claims simplicity ("simple"); no public code found

### [Enriching Sequential Recommendation with Graph Laplacian Positional Embeddings](https://arxiv.org/abs/2609.31253)
`2609.31253` · score 40.0 · analysis: heuristic · feasibility 5/10

Sequential recommenders typically rely on learnable positional embeddings to encode the order of user interactions. In this work, we ask whether this ordinal signal can be replaced by a structural one derived from the item space.

Rubric evidence: + public benchmarks ("benchmarks"); + claims simplicity ("simple"); no public code found; + keywords: ranking

### [Embedding Subspace Partitioning for Dynamic Multi-Objective Retrieval](https://arxiv.org/abs/2609.30601)
`2609.30601` · score 37.9 · analysis: heuristic · feasibility 5/10

Modern industrial recommender systems must optimize across competing objectives, balancing semantic relevance with business metrics such as engagement and revenue. While bi-encoders dominate large-scale retrieval due to their efficiency, they collapse these heterogeneous signals into a single static embedding space.

Rubric evidence: + public benchmarks ("benchmark"); + formal statement ("bound"); - system/framework (many moving parts) ("framework"); no public code found; + keywords: retrieval

### [SignTrace: Describe a Sign, Find the Word](https://arxiv.org/abs/2609.30295)
`2609.30295` · score 33.7 · analysis: heuristic · feasibility 4/10

Identifying an unfamiliar sign is difficult when a learner remembers its movement but does not know its meaning or formal feature codes. SignTrace addresses this longstanding reverse-lookup problem through natural-language access to a Chinese sign-language dictionary.

Rubric evidence: - LLM-based ("LLM"); + public benchmarks ("benchmark"); no public code found; + keywords: retrieval, ranking

### [Recommendation World Models for Future-State Control](https://arxiv.org/abs/2609.30711)
`2609.30711` · score 32.6 · analysis: heuristic · feasibility 4/10

Sequential recommendation optimizes which items to rank, while each displayed slate also shapes subsequent feedback and user state. We study how a trained ranker can support decisions about these future consequences.

Claimed: Recall@20 20.0 (abstract (regex extraction, unverified))
Rubric evidence: no public code found

### [SPADE: Escaping the Popularity-Similarity Frontier to Measure Serendipitous Recommendations](https://arxiv.org/abs/2609.31164)
`2609.31164` · score 32.6 · analysis: heuristic · feasibility 4/10

Recommender systems engineer serendipity to foster active exploration and break predictable consumption cycles. The problem with existing offline beyond-accuracy metrics is that they often either isolate historical similarity or global popularity.

Rubric evidence: no public code found

### [Retail Product Search: A Practical Approach at Target](https://arxiv.org/abs/2609.31498)
`2609.31498` · score 32.6 · analysis: heuristic · feasibility 4/10

Search is one of the most important features in e-commerce, directly driving customer engagement and business growth. A good product search system must show both relevant and desirable results.

Rubric evidence: no public code found

### [REALMS: An AI-Assistant Conversational System for Real-Time Exact Audience Sizing over High-Dimensional Nested Profiles](https://arxiv.org/abs/2609.30547)
`2609.30547` · score 32.6 · analysis: heuristic · feasibility 4/10

Audience sizing is a critical component of digital marketing. It enables precise resource allocation, campaign planning, and performance optimization.

Rubric evidence: - LLM-based ("LLM"); no public code found; + keywords: sampling, retrieval

## All screened papers

| score | paper | compute | data | clarity | code | interest | analysis | laptop-feasible |
|---:|---|---:|---:|---:|---:|---:|---|---|
| 45.8 | [Bootstrapping Conversational Recommendation Agents At Spotify: Synthetic Data Generation and Self-Improvement Loops](https://arxiv.org/abs/2609.30297) `2609.30297` | 0.65 | 0.75 | 0.45 | 0.00 | 0.00 | heuristic | True |
| 42.1 | [RecToolBench: Benchmarking Recommendation-Specific Tool Orchestration under Fuzzy User Intent](https://arxiv.org/abs/2609.30717) `2609.30717` | 0.30 | 0.45 | 0.60 | 1.00 | 0.00 | heuristic | False |
| 41.6 | [CG-Probes: Recovering Guardrail Directions from Patient Query Embeddings](https://arxiv.org/abs/2609.31062) `2609.31062` | 0.45 | 0.75 | 0.55 | 0.00 | 0.00 | heuristic | False |
| 40.0 | [Enriching Sequential Recommendation with Graph Laplacian Positional Embeddings](https://arxiv.org/abs/2609.31253) `2609.31253` | 0.50 | 0.45 | 0.55 | 0.00 | 0.20 | heuristic | True |
| 37.9 | [Embedding Subspace Partitioning for Dynamic Multi-Objective Retrieval](https://arxiv.org/abs/2609.30601) `2609.30601` | 0.50 | 0.45 | 0.45 | 0.00 | 0.20 | heuristic | True |
| 33.7 | [SignTrace: Describe a Sign, Find the Word](https://arxiv.org/abs/2609.30295) `2609.30295` | 0.30 | 0.45 | 0.40 | 0.00 | 0.40 | heuristic | False |
| 32.6 | [Recommendation World Models for Future-State Control](https://arxiv.org/abs/2609.30711) `2609.30711` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | heuristic | True |
| 32.6 | [SPADE: Escaping the Popularity-Similarity Frontier to Measure Serendipitous Recommendations](https://arxiv.org/abs/2609.31164) `2609.31164` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | heuristic | True |
| 32.6 | [Retail Product Search: A Practical Approach at Target](https://arxiv.org/abs/2609.31498) `2609.31498` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | heuristic | True |
| 32.6 | [REALMS: An AI-Assistant Conversational System for Real-Time Exact Audience Sizing over High-Dimensional Nested Profiles](https://arxiv.org/abs/2609.30547) `2609.30547` | 0.30 | 0.40 | 0.40 | 0.00 | 0.40 | heuristic | False |
| 31.6 | [QReason: Query-Focused Decoupled Chain-of-Thought for Efficient Passage Reranking](https://arxiv.org/abs/2609.30904) `2609.30904` | 0.30 | 0.45 | 0.30 | 0.00 | 0.40 | — | — |
| 30.5 | [Nearest but Not Dearest: Shared Curator-Feedback Infrastructure for Content-Only Search and Recommendation](https://arxiv.org/abs/2609.30568) `2609.30568` | 0.50 | 0.00 | 0.55 | 0.00 | 0.20 | — | — |
| 30.5 | [T-RoPE: Time-Aware Rotary Position Embedding for Sequential Recommendation](https://arxiv.org/abs/2609.30576) `2609.30576` | 0.30 | 0.45 | 0.55 | 0.00 | 0.00 | — | — |
| 29.5 | [Where Does Retrieval-Based Open-Ended Evaluation Fail? Automatic Taxonomy Induction from Long-Form Medical Answer Factuality Verification](https://arxiv.org/abs/2609.30467) `2609.30467` | 0.30 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 26.3 | [AgentRecommender: LLM Agents Enable Customizable Recommender Systems on the User Side](https://arxiv.org/abs/2609.31166) `2609.31166` | 0.30 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 24.2 | [AutoResearch at Production Scale: Failure Modes and a Multi-Agent Framework](https://arxiv.org/abs/2609.30541) `2609.30541` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 22.1 | [Component Benchmark: Hierarchical Model Profiling for Large-scale Recommendation Systems](https://arxiv.org/abs/2609.30656) `2609.30656` | 0.00 | 0.45 | 0.60 | 0.00 | 0.00 | — | — |
| 17.9 | [KuaFu: Compressing Long User Behavior into Understanding at Billion Scale](https://arxiv.org/abs/2609.31045) `2609.31045` | 0.00 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 17.9 | [Epstein Files Engine: Agentic Search for Investigative Journalism](https://arxiv.org/abs/2609.30611) `2609.30611` | 0.30 | 0.40 | 0.00 | 0.00 | 0.00 | — | — |

Scores: weighted mean of rubric criteria (0–1 each) × 100; see README for the rubric.