Deep Dives
In-depth AI analysis, industry observations, and technical insights — slow down and read.

LLM Reinforcement Learning: Is the Importance Sampling Ratio Really Necessary? LGPO Says No
New research proposes LGPO, questioning the need for importance sampling ratios in LLM RL. By replacing trust region constraints with likelihood gating, it eliminates old-policy likelihood computation, reducing training time by ~18% on average while maintaining stability, and remains stable even in fully on-policy scenarios where GRPO fails.

AraCrypto-ML: A Multi-label Corpus of Arabic Cryptocurrency Discourse
AraCrypto-ML is the first multi-label corpus of Arabic Facebook cryptocurrency discourse, containing 10,000 posts annotated with six communicative functions, with model annotation agreement surpassing human annotators and revealing differences in engagement patterns across post functions.

SimRec: Reinforcement Learning Reshapes the Search and Follow-up Strategies of Recommendation Agents
SimRec uses reinforcement learning to train the complete interaction strategy of a recommendation agent, with a task-anchored user simulator and task-aligned rewards significantly improving NDCG on ESCI and Amazon C4, letting a 4B model surpass larger training-free agents.

Making LLMs Dumber? An Empirical Study of Performance Throttling and Believable Capability Reduction
An empirical study of believable LLM performance throttling: weak-teacher LoRA precisely controls scores but creates token attractors in multiple choice, with error distributions differing from real weak models.

QDRAG: How Query Decomposition Improves RAG Citation Faithfulness and Generalization
The paper proposes QDRAG, a framework that decomposes questions into atomic sub-questions and uses re-ranking pruning to eliminate over-citation, significantly improving RAG citation faithfulness and generalization, surpassing Gemini-Pro on multimodal QA benchmarks.

AutoOntic: Automatically Inducing Ontologies from Text to Build High-Quality Knowledge Graphs
AutoOntic is a pipeline that automatically induces ontologies from text. By canonicalizing entity types and relations, inducing hierarchies, and mining domain/range constraints, it significantly improves knowledge graph quality. On Text2KGBench, over 90% of types semantically align with reference ontologies, and retrieval accuracy on the MINE dataset reaches 88%.

DecArena: A Large-Scale Benchmark for Decentralized Multi-Agent Collaboration
DecArena introduces a large-scale benchmark for decentralized multi-agent collaboration under graph-constrained communication, with distributed tool-calling tasks over 100K tools and a GraphHarness baseline: +48.5 task success, +30.6 collaboration quality, +34.0 cost improvement.

Learning Calibrated Decision Heads from Bandit Feedback: CAPS Wins
A study fine-tunes decision heads on Qwen3 via LoRA from bandit feedback: naive REINFORCE rewards collapse performance, while CAPS beats confirmed-answer-only fine-tuning but not a cross-entropy baseline under full labels.

EASTAR: Event-Anchored and Task-Adaptive Urban Region Representation Learning
EASTAR proposes an event-anchored and task-adaptive urban region representation learning framework that decomposes mobility representations using offline activity events as semantic anchors and designs a task-adaptive prompt router with LLM priors, reducing MAE by up to 39.15% across four prediction tasks.

Accurate Predictions ≠ Correct Causality: The Causal Response Dilemma of Temporal Foundation Models
The paper evaluates the causal response capability of tabular and time series foundation models, finding that forecast accuracy does not indicate price effect estimation precision; more context improves forecasting but not causal estimation, and instrumental variables are effective only in specific forms.

AI4Plasma: Five Surrogate Models Compete on Data Efficiency, FNO Wins
AI4Plasma compares five surrogate models for CCP field prediction data efficiency: FNO needs the fewest training cases in most scenarios, but the optimal architecture depends on accuracy targets and parameter regions.

A New Approach to Federated Class Incremental Learning: Using Pseudo Features to Overcome Catastrophic Forgetting
This article introduces FCLPF, a federated class incremental learning method that uses prototypes to generate pseudo features to mitigate catastrophic forgetting without storing old samples or using GANs. On CIFAR-100 and TinyImageNet, FCLPF significantly outperforms baselines and greatly reduces forgetting, offering a lightweight solution for federated learning in dynamic data scenarios.

No Training Needed! MPPA Breaks Diffusion LLM Long-Context Bottleneck with Mask Patterns
Diffusion LLMs struggle with long context due to RoPE limits. MPPA, a training-free extrapolation method, uses mask density to dynamically adjust position encoding, cutting LLaDA-8B perplexity at 8192 tokens from 19.51 to 3.12 with 100% retrieval accuracy.

Hacker News Hot Debate: Why Believe $100K a Year Is Fair?
A Hacker News post sparks debate on whether $100K is a fair tech salary. This analysis explores regional differences, skill value, and market supply and demand.

CTF Red Teaming Reveals: Non-Cybersecurity Models Are Better at Finding Sandbox Vulnerabilities
SPACE open-source sandbox platform releases new red team findings: CTF tests need explicit prompts to break host-guest isolation; non-cybersecurity models outperform security-specialized models at finding sandbox escape vulnerabilities; the platform is robust against isolation escapes.

Shipping Through Rain or Snow: The Hidden Engineering Culture Code of AI Teams
Starting from a "rain or snow, we keep shipping" tweet, this article analyzes the real meaning of continuous delivery culture in AI teams, engineering resilience requirements, and external communication strategies, with three actionable takeaways for practitioners.

Jensen Huang's 33-Year Leadership Praised: The AI Legend from Startup to America's Most Valuable Company
A tweet pays tribute to Jensen Huang's 33 years of leadership: from startup to America's most valuable company, his relentless improvement philosophy, generous founder mentorship, and impact on AI.

Perplexity Teammates: Letting an AI Team Produce Product Marketing Videos End to End
Perplexity Teammates enables an AI team to create product marketing videos end to end, covering scripts, assets, and final cuts. This article analyzes the potential impact of multi-agent collaboration on marketing content production and the issues still to be verified.

Perplexity Mac App Hybrid Mode: Cloud Orchestrator + Local Subagents
Perplexity's Mac app introduces hybrid mode: a cloud orchestrator handles scheduling while local subagents execute tasks. Usage keeps growing, per the official tweet. We analyze the architecture and why it matters.

Decider v1.3 Open-Sourced: New 'locate' Feature Added, Decision Model Performance Significantly Improved
Decider v1.3 and Decider Flash v1.3 are open-sourced with a new locate feature supporting text localization and character offsets. Significantly outperforms general extraction models on 17 out-of-domain tasks, with a 6.54 percentage point accuracy gain on production classification. Weights are public on Hugging Face, free on OpenRouter for a limited time.