Daily Digest — Oct 6

BMFM-RNA: whole-cell expression decoding improves transcriptomic foundation models

• Whole-cell expression decoding (WCED) outperforms masked language modeling (MLM) in transcriptomic foundation models for downstream tasks, despite higher training reconstruction error. • WCED reconstructs the entire gene vocabulary from a single CLS token embedding, enabling improved cell representations by creating a maximally informative bottleneck.

Better Retrieval, Limited Clustering Gains: A Controlled Study of Multilingual Company Entity Resolution

• A multi-view selector increases direct pair recall from 53.74% to 76.98% at 72,660 candidate edges, yet cluster recall only rises marginally from 32.54% to 33.33%. • The evaluation on a GLEIF sample of 3,633 names and 2,880 source identities shows that 202 out of 208 newly retrieved silver-positive pairs fall below the decision threshold.

Aligning Multimodal Patient Evidence with Biomedical Knowledge Graphs for Clinical LLMs

• The MM-KG framework explicitly links multimodal patient observations with biomedical knowledge graphs using route-prioritized alignment, yielding drug-controlled AUROC interaction improvements of +0.194 on MIMIC-IV and +0.299 on ADNI datasets.

Wikidata Search Traces: A Dataset for Training Knowledge Graph Search Agents

• Language model agents improve knowledge graph search accuracy by utilizing a recursive language model harness that batches graph calls and manages retrieved evidence in persistent Python state rather than direct tool calling. • The researchers construct multi-hop questions on a frozen Wikidata snapshot and release 10,235 solving traces alongside the harness to train graph search agents.

Curriculum Brain: Constructing Curriculum Knowledge Graphs as a Substrate for Cognitive Diagnosis

• Curriculum Brain automates the construction of curriculum knowledge graphs by pairing a version-controlled knowledge base with an agentic pipeline of eleven single-responsibility agents. • The system processes official curriculum documents and textbooks to generate and validate concept-skill mappings, achieving a 91.5% resolution rate without human escalation in later pipeline configurations.

EgoSelf: From Memory to Personalized Egocentric Assistant

• EgoSelf introduces a graph-based interaction memory that captures temporal and semantic relationships from past observations to construct user-specific profiles for personalized egocentric assistants.

Why Better Cross-Lingual Alignment Fails for Better Cross-Lingual Transfer: Case of Encoders

• Explicit cross-lingual alignment techniques fail to improve token-level downstream performance because alignment and downstream task objectives are largely orthogonal. • Analysis of four XLM-R encoder models shows embedding distances are unreliable predictors of task performance improvements, and alignment and task gradients often conflict.

Meta-Reinforcement Learning with Self-Reflection for Agentic Search

• Meta-Reinforcement Learning with Self-Reflection (MR-Search) trains agents to adapt search strategies across episodes by generating explicit self-reflections. • This approach improves in-context exploration at test-time by leveraging self-reflections as additional context for subsequent search attempts.