arXiv.org

1746 articles in the Truth Foundry index. Each links out to the original.

• Debiasing NLP Models Causes Unintended Stereotyping of Other Groups · Preprocessing-based stereotype mitigation induces counter-stereotyping side effects for unrelated demographics.
• Persona Steering Degrades AI Performance in Educational Settings · Activation-based persona steering significantly lowers AI answer quality and alters scoring in educational applications.
• CausalDS Benchmark Evaluates Causal Reasoning in Data-Science Agents · CausalDS benchmark tests LLMs on causal reasoning, data analysis, and tool use in synthetic data-science workflows.
• Gemini Models Show High Reliability as Audio Judges for Voice Agents · Gemini 2.5 Flash audio judges match human ratings on 7 of 8 dimensions for full-duplex voice agent conversations.
• Hallucination Self-Play Framework Bootstraps LLM Detectors via Evolved Generators · Hallucination Self-Play enables detectors to bootstrap with evolved generators for improved faithfulness detection.
• New MiniLM Method Achieves State-of-the-Art Out-of-Scope Intent Detection · A multi-cluster boundary learning method using MiniLM embeddings detects out-of-scope intents with state-of-the-art accuracy.
• Barenholtz's Autogenerative Theory Enriches Harris's Integrationist Linguistics · Elan Barenholtz's autogenerative theory fills explanatory gaps in Roy Harris's Integrationist linguistics regarding LLMs.
• DeepSearch-Evolve Framework Enables Self-Improving Web Agents · DeepSearch-Evolve uses a verifiable environment to train self-improving web agents without external distillation.
• TACO Method Mitigates Credit Contamination in LLM Reinforcement Learning · TACO calibrates credit assignment to suppress reinforcement of erroneous tail tokens in LLMs.
• AI4Math Shifts from Problem-Solvers to Frontier Research Agents · New position paper argues AI must evolve from theorem solvers to autonomous mathematical research agents.
• LSTM Networks Outperform Traditional Models in Twitter Sentiment Analysis · LSTM networks achieve superior accuracy over traditional ML models for Twitter sentiment classification
• Researchers Create Spanish Stereotype Dataset to Test LLM Cultural Bias · New EspanStereo dataset uses human-LLM collaboration to benchmark Spanish-speaking LLMs for culturally specific biases.
• Self-Evolving LLM Agents Reduce Latency and Errors in Industrial Systems · Tool-making pipeline compiles SOPs into versioned tools, cutting latency by 42% and errors by 53% in production alarms.
• Structured Pruning of LLMs via Power Transformation and Sign-Preserving Score Aggregation · New method adapts Adaptive Feature Retention to structured pruning for LLMs
• LEXIC Model Boosts Eye-Movement Reading Prediction Without Language Models · LEXIC-Base injects word-level difficulty signals into gaze-only models to improve reading comprehension prediction.
• Probing Internal Representations Improves LLM Forecaster Calibration · Internal representation probes outperform chain-of-thought for calibrating LLM forecasters
• Researchers Introduce GRAPHEVAL Framework to Quantify LLM Reasoning Uncertainty · A new graph-based framework quantifies LLM reasoning fidelity and exposes flaws in Self-Consistency decoding.
• New Framework Reveals LLMs Struggle to Lower Cognitive Demand in Education · Research shows LLMs reliably increase task difficulty but fail to simplify for learners.
• Hol-PCFG Achieves SOTA Unsupervised Parsing with 99.94% Parameter Reduction · Holographic Neural PCFG (Hol-PCFG) recasts grammar scoring as algebraic relations, achieving state-of-the-art performance across six languages with minimal parameters.
• Researchers Release PLURAL Dataset for Culturally Diverse AI Alignment · New dataset PLURAL uses global survey data to align AI with diverse cultural values.
• COALA Framework Enhances Speech Recognition via Contextual Biasing · COALA improves ASR systems by mapping latent representations to quantify audio-entity matching intensity.
• MASTE: Multi-Agent Pipeline Enables Zero-Shot Aspect Sentiment Triplet Extraction · MASTE uses a multi-agent pipeline to achieve superior zero-shot ASTE performance without training data.
• COBART Model Boosts Ad Headline CTR and Quality by 5.82% and 25.82% Respectively · COBART uses prefix control tokens and BART fine-tuning to optimize ad headlines for higher CTR and quality.
• Researchers Achieve 17.97% Error Rate in Multilingual Speech Recognition Challenge · A new diarization-guided ASR system significantly improves multilingual speech recognition accuracy on the MLC-SLM 2026 Challenge.
• Researchers Release Grounded 8-K Event Taxonomy with LLM Validation · New system tags 601k SEC 8-K filings with precise event types using LLMs and fuzzy matching.
• ICDAR 2026 Competition Evaluates LLMs for Historical Document OCR Correction · Four teams tested LLM-based post-correction on noisy OCR of 17th-20th century historical documents.
• SQuaD-SQL Enables Small Language Models to Match LLM Performance on Text-to-SQL · SQuaD-SQL uses knowledge distillation to let small language models perform Text-to-SQL tasks efficiently.
• XAlpha: A Memory-Driven AI Quant Researcher for Alpha Discovery · XAlpha is a memory-driven AI system that automates end-to-end quantitative research and alpha discovery in financial markets.
• Cross-Family Verifier Ensembles Improve Text-to-Speech Evaluation · Using diverse ASR evaluators reduces bias in Best-of-N TTS selection
• Hidden Decoding Enables Sequence-Length Scaling for Frontier LLMs · Hidden Decoding expands token computation via sequence-length scaling for fixed-backbone LLMs