-
[Paper Review] MobileExplorer: Accelerating On-Device Inference for Mobile GUI Agents via Online Exploration
Instead of letting slow on-device VLM inference idle, MobileExplorer spends that window probing the screen to gather hints for the next reasoning step.
-
[Paper Review] Ferret-UI Lite: Lessons from Building Small On-Device GUI Agents
Apple's 3B on-device GUI agent. A close read of how real+synthetic data curation, zoom-in visual tool-use, and a two-stage SFT→RLVR recipe push a small model as far as it can go on GUI grounding and navigation — and where it honestly hits a wall.
-
[Paper Review] ShowUI: One Vision-Language-Action Model for GUI Visual Agent
A GUI agent that 'sees' screenshots like a human and clicks. A deep dive into ShowUI's UI-guided token selection and interleaved vision-language-action streaming, which hit 75.1% zero-shot grounding with a 2B model and 256K data.
-
[Paper Review] MinT: Managed Infrastructure for Training and Serving Millions of LLMs
A managed infrastructure for training and serving millions of LoRA policies over a small set of resident base models. Adapter revisions become the unit that crosses the training-serving boundary, cutting the handoff by 18.3x and validating the path up to 1T-scale MoE.
-
[Paper Review] Goldfish: Monolingual Language Models for 350 Languages
1,154 small monolingual GPT-2 models covering 350 low-resource languages. With 5MB-1GB of byte-premium-scaled text and 39M-125M parameters, they consistently beat XGLM 4.5B, BLOOM 7.1B, and MaLA-500 10B on FLORES perplexity.
-
[Paper Review] Algorithmically Establishing Trust in Evaluators
A zero-knowledge-style challenge–response protocol that certifies the trustworthiness of an evaluator (e.g. LLM-as-a-judge) without any labelled data, bounded by (1/4)^r.
-
[Paper Review] RTP-LX: Can LLMs Evaluate Toxicity in Multilingual Scenarios?
AAAI-25 paper introducing RTP-LX: a human-transcreated, human-annotated corpus of 1,100 toxic prompts across 28 languages, used to stress-test 10 S/LLMs as multilingual safety evaluators.
-
[Paper Review] Improving Large-Scale Recommender Systems with Auxiliary Learning
Analysis of C2AL, Meta's framework that combats majority-cohort bias in large-scale ads recommendation models via cohort-contrastive auxiliary learning.
-
[Paper Review] AgenticRecTune: Multi-Agent with Self-Evolving Skillhub for Recommendation System Optimization
An LLM multi-agent framework that automates the tuning of system-level configuration (fusion weights, demotion weights, diversity thresholds) across all three stages — pre-ranking, ranking, re-ranking — of Google Discover. Five specialized agents (Actor, Critic, Insight, Skill, Online) form a closed loop that feeds live A/B results back into memory and a self-evolving skillhub, simultaneously lifting engagement and diversity in production.
-
[Paper Review] Graph-Based Audience Expansion Model for Marketing Campaigns
A two-stage lookalike model from Rakuten — TransE-style pretraining on a 70+ service cross-service knowledge graph, then a GCN that aggregates 'knowledge queries' (head + relation embeddings) rather than neighbor entities. SIGIR 2024 short paper.
