Weekly Data/ML Paper Briefing

Date: 2026-02-16 (Week of Feb 10-16)

---

Report 1: LLM & AI Technology

1.1 RAGLens: Faithful RAG with Sparse Autoencoders

Item	Details
Title	Toward Faithful Retrieval-Augmented Generation with Sparse Autoencoders
Authors	Guangzhi Xiong et al.
Venue	ICLR 2026
Link	arXiv:2512.08892

Core Idea: RAG systems suffer from "faithfulness failures" where outputs contradict or extend beyond retrieved sources. RAGLens uses Sparse Autoencoders (SAEs) to disentangle LLM internal activations and identify features specifically triggered during hallucinations.

Key Contributions:

• Information-based feature selection pipeline for hallucination detection
• Lightweight detector using LLM internal representations (no external LLM judges needed)
• Provides interpretable rationales for decisions
• Enables post-hoc mitigation of unfaithful RAG outputs

Practical Applicability: Directly applicable to the vacant house chatbot project. Can integrate RAGLens to detect when Text-to-SQL or policy document RAG outputs hallucinate, improving response reliability.

---

1.2 Aletheia: Autonomous Math Research Agent

Item	Details
Title	Towards Autonomous Mathematics Research
Authors	Google DeepMind
Release	Feb 2026
Link	GitHub/DeepMind

Core Idea: Aletheia bridges competition-level math solving to professional research. Uses an agentic loop: Generator proposes solutions, Verifier checks for flaws, Reviser corrects errors.

Key Findings:

• 100x compute reduction vs 2025 version for IMO-level problems
• 95.1% accuracy on IMO-Proof Bench Advanced (vs previous 65.7%)
• Autonomously generated publishable research paper (Feng26)
• Resolved 4 open Erdos Conjectures without human intervention

Architecture:

+-------------+     +------------+     +-----------+
|  Generator  | --> |  Verifier  | --> |  Reviser  |
| (Candidate) |     | (Check)    |     | (Correct) |
+-------------+     +------------+     +-----------+
       ^                                     |
       +-------------------------------------+
              (Iterate until approved)

Practical Applicability: The Generator-Verifier-Reviser pattern is applicable to complex query validation in Text-to-SQL systems. Can implement similar verification loops for SQL query generation.

---

1.3 Agentic RAG Survey

Item	Details
Title	Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG
Authors	Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei
Venue	CoRR 2025
Link	dblp

Core Idea: Comprehensive survey on combining RAG with agentic capabilities - systems that can reason, act via tools, and interact in goal-directed ways.

Key Themes:

• RAG + Tool Use integration patterns
• Multi-step planning with retrieval
• Self-correction mechanisms in RAG agents
• Evaluation frameworks for agentic RAG

Practical Applicability: Reference material for evolving the vacant house chatbot from simple RAG to agentic RAG with multi-step reasoning and tool orchestration.

---

Report 2: Forecasting Models & Data Science

2.1 SE-LLM: Semantic-Enhanced Time Series Forecasting

Item	Details
Title	Semantic-Enhanced Time-Series Forecasting via Large Language Models
Authors	Hao Liu et al.
Venue	arXiv
Link	arXiv:2508.07697

Core Idea: SE-LLM bridges the modality gap between linguistic knowledge and time series patterns by embedding periodicity and anomaly characteristics into semantic space.

Key Innovations:

• Explores inherent periodicity and anomalies of time series
• Embeds temporal characteristics into semantic space for LLM understanding
• Plugin module for self-attention modeling both long-term and short-term dependencies
• Freezes LLM weights, reducing computational cost

Architecture:

Time Series --> [Periodicity Extractor] --> Token Embedding
     |                                           |
     +--> [Anomaly Detector] ------------------> |
                                                 v
                              [Frozen LLM + Plugin Module]
                                                 |
                                                 v
                                            Forecast

Practical Applicability: For vacant house prediction: can leverage LLM understanding of semantic patterns (e.g., "housing market decline," "population aging") combined with numerical time series for improved forecasting.

---

2.2 DMamba: Decomposition-Enhanced Mamba for Time Series

Item	Details
Title	DMamba: Decomposition-enhanced Mamba for Time Series Forecasting
Authors	DMambaKDD Team
Venue	KDD 2026 (expected)
Link	arXiv:2602.09081

Core Idea: State Space Models (Mamba) struggle with non-stationary data. DMamba decomposes time series into trend and seasonal components, processing each with appropriately complex models.

Key Insight:

Component	Nature	Recommended Model
Trend	Low-dimensional, linear constraints	Simple MLP
Seasonal	High-dimensional, dynamic interactions	Mamba encoder

Architecture:

Input --> [EMA Decomposition] --> Trend --> [MLP] --------+
                |                                         |
                +--> Seasonal --> [Mamba Encoder] --------+--> Forecast

Results:

• New SOTA on ETT, Weather, PEMS benchmarks
• Outperforms both Mamba-based and decomposition-based models
• O(L) complexity vs O(L^2) for Transformers

Practical Applicability: Excellent fit for vacant house prediction pipeline. Housing trends are low-dimensional (driven by macro factors), while seasonal patterns (moving seasons, fiscal year effects) are complex. DMamba's architecture matches this structure.

---

2.3 Deep Learning for Electricity Price Forecasting

Item	Details
Title	Deep Learning for Electricity Price Forecasting: A Review
Authors	Multiple (review paper)
Link	arXiv:2602.10071

Key Findings:

• Two-phase approach: Prefill (compute attention states) + Decode (autoregressive generation)
• Hybrid models combining linear methods with neural networks show robust performance
• Importance of proper evaluation beyond RMSE/MAE (economic metrics)

Practical Applicability: Methodology insights applicable to any time series forecasting task, including infrastructure-related predictions like housing market dynamics.

---

Report 3: Spatial Data Analysis

3.1 AURA: Autonomous AI for Affordable Housing Site Selection

Item	Details
Title	Autonomous AI Agents for Real-Time Affordable Housing Site Selection
Authors	Olaf Y. L. Imanov et al. (DTU, Turkey, Latvia)
Venue	arXiv
Link	arXiv:2602.03940

Core Idea: AURA (Autonomous Urban Resource Allocator) uses multi-agent reinforcement learning for real-time affordable housing site selection under 127 regulatory constraints.

System Architecture:

+------------------------+
|   Coordination Agent   |
+------------------------+
     |    |    |    |
     v    v    v    v
+------+ +------+ +------+ +------+
|Geo-  | |Regu- | |Multi-| |Consen|
|spatial| |latory| |Obj.  | |sus   |
|Agent | |Agent | |Agent | |Agent |
+------+ +------+ +------+ +------+
   |        |        |
   v        v        v
[GNN]  [Constraint] [Pareto]
       [Reasoning]  [Policy]

Results:

Metric	Value
Regulatory Compliance	94.3%
Pareto Hypervolume Improvement	+37.2%
Site Selection Time	72 hours (vs 18 months)
Transit Accessibility	+31% vs human experts
Environmental Impact	-19% vs human experts

Practical Applicability: Highly relevant to vacant house prediction project. The multi-agent approach with regulatory compliance verification can be adapted for:

• Identifying high-risk vacant areas under zoning constraints
• Optimizing resource allocation for vacant house management
• Multi-objective optimization (cost, accessibility, social equity)

---

3.2 Geospatial ML for CVD Mortality Analysis

Item	Details
Title	Geospatial and ML analyses of CVD mortality across the US
Authors	Multiple (BMC Public Health)
Venue	BMC Public Health, Feb 2026
Link	Springer

Methods Used:

Technique	Purpose
Global Moran's I	Spatial autocorrelation detection
Getis-Ord Gi*	Hot spot / clustering analysis
Shapley Values	Feature importance interpretation
ML Models	Mortality prediction

Key Insight: Combining spatial clustering (Moran's I, Getis-Ord) with ML interpretability (SHAP) provides both predictive power and actionable insights.

Practical Applicability: The Moran's I + Getis-Ord + SHAP combination is directly applicable to vacant house clustering analysis. Already using similar methods in the prediction pipeline (Step 3: clustering). Can add SHAP for better interpretability.

---

3.3 Interpretable Property Market Models

Item	Details
Title	Modern approaches to building interpretable models of the property market
Authors	Multiple
Link	arXiv:2506.15723

Focus: Using ML for mass cadastral valuation with interpretability focus.

Key Methods:

• Ensemble methods with feature importance
• Spatial relationships in property pricing
• Highlighting significant factors in price formation

Practical Applicability: Reference for adding interpretability to vacant house prediction models. Important for stakeholder communication and policy recommendations.

---

Summary Table

Category	Paper	Key Tech	Relevance
LLM/RAG	RAGLens	SAE, Hallucination Detection	Chatbot reliability
LLM/Agent	Aletheia	Generator-Verifier-Reviser	Query validation
Forecasting	SE-LLM	LLM + Time Series	Semantic forecasting
Forecasting	DMamba	Mamba + Decomposition	SOTA time series
Spatial	AURA	Multi-agent RL	Site selection
Spatial	CVD Mortality	Moran's I + SHAP	Spatial ML

---

Action Items for Current Projects

1. Vacant House Prediction Pipeline:
2. Consider DMamba architecture for Step 4 (model training)

3. Add SHAP values for model interpretability

4. Vacant House Chatbot:

5. Implement RAGLens-style verification for RAG outputs

6. Consider Generator-Verifier pattern for Text-to-SQL

7. Documentation:

8. Add AURA paper to architecture references (multi-agent spatial optimization)
9. Update spatial analysis section with Moran's I + Getis-Ord methodology

---

Generated: 2026-02-16 09:00 KST