Tech Business/VC Hacker News

Why eval startups fail (2025)

An analysis of why startups focusing on LLM evaluation (evals) often fail, likely discussing the difficulty of creating reliable benchmarks.

AI/ML arXiv cs.AI

LLMs Prompted for Legal Context Object More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context

Research showing that small on-premises LLMs are prone to 'overrefusal' when prompted with authority-style legal prefixes, potentially introducing bias.

Cybersecurity arXiv cs.AI

AdversaBench: Automated LLM Red-Teaming with Multi-Judge Confirmation and Cross-Model Transferability

Introduction of AdversaBench, an automated red-teaming pipeline for LLMs that uses multi-judge confirmation to identify real failures.

AI/ML arXiv cs.AI

ASALT: Adaptive State Alignment for Lateral Transfer in Multi-agent Reinforcement Learning

Presentation of ASALT, a method for multi-agent reinforcement learning that enables knowledge transfer between domains with mismatched state-space dimensions.

AI/ML arXiv cs.AI

Uncertainty-Aware Longitudinal Forecasting of Alzheimer's Disease Progression Using Deep Learning

A deep learning framework for forecasting Alzheimer's progression using Temporal Fusion Transformers and uncertainty estimation to improve clinical reliability.

AI/ML arXiv cs.AI

ScaleToT: Generalizing Structured LLM Reasoning for Billion-Scale Low-Activity User Modeling

ScaleToT introduces a method to generalize structured LLM reasoning to billions of low-activity users by training lightweight student models on a small high-quality subset.

AI/ML arXiv cs.AI

AI Tokenomics: The Economics of Tokens, Computation, and Pricing in Foundation Models

A theoretical framework for 'AI Tokenomics,' examining the economic relationship between tokens, computation, energy, and pricing in foundation models.

Software Engineering arXiv cs.AI

Abstractions of Queries in Ontology-Based Data Access

A study on query abstraction in ontology-based data access, specifically focusing on minimally complete and maximally sound abstractions using existential rules.

Software Engineering arXiv cs.AI

When CQs Go Wrong: Challenges in CQ Verification with OE-Assist

An investigation into the challenges of Competency Question (CQ) verification in ontology engineering and the role of LLM assistants in refining those questions.

Tech Business/VC Hacker News

Ashby (YC W19) Is Hiring EMEA Engineers Who Can Design

Ashby is looking for engineers in the EMEA region who possess strong design skills.

Software Engineering Hacker News

Show HN: Graphical SQL Builder and Debugger

A new graphical SQL builder and debugger has been showcased on Hacker News to simplify database query construction and troubleshooting.

AI/ML arXiv cs.AI

CompressKV: Semantic-Retrieval-Guided KV-Cache Compression for Resource-Efficient Long-Context LLM Inference

CompressKV is a new framework for GQA-based LLMs that optimizes KV-cache compression by identifying Semantic Retrieval Heads to maintain performance with minimal memory usage.

AI/ML arXiv cs.AI

The Latent Bridge: A Continuous Slow-Fast Channel for Real-Time Game Agents

Researchers introduced the Latent Bridge, a continuous channel connecting a slow reasoning VLM and a fast reactive VLM for real-time game agent control.

AI/ML arXiv cs.AI

On the Smallness of the Large Language Models Scaling Exponents

This paper examines the sustainability of LLM scaling exponents, arguing that current energy resource trajectories are unsustainable despite potential numerical biases.

AI/ML arXiv cs.AI

A specialized reasoning large language model for accelerating rare disease diagnosis: a randomized AI physician assistance trial

RaDaR is an open-source 32B reasoning LLM specifically trained to assist physicians in the diagnosis of rare diseases, showing significant accuracy improvements.

AI/ML arXiv cs.AI

Reinforcement Learning for Computer-Use Agents with Autonomous Evaluation

A new RL fine-tuning framework uses autonomous vision-language evaluation as a reward signal to improve the performance of Computer-Use Agents in GUI environments.

AI/ML arXiv cs.AI

Governed Shared Memory for Multi-Agent LLM Systems

The paper introduces MemClaw, a production multi-tenant memory service for multi-agent LLM systems designed to prevent leakage and provenance collapse.

AI/ML arXiv cs.AI

GUI vs. CLI: Execution Bottlenecks in Screen-Only and Skill-Mediated Computer-Use Agents

A benchmark comparing GUI and CLI agents reveals that while GUI agents have higher raw success, CLI agents are more effective when provided with comprehensive skill interfaces.

AI/ML arXiv cs.AI

Quant Convergence: Bridging Classical Value Investing and Modern Factor Models for Systematic Equity Selection

Research demonstrates that integrating classical Benjamin Graham value investing rules into modern AI equity models reduces volatility and improves risk-adjusted returns.

Software Engineering Hacker News

Lithp.py (~2008)

A mention of Lithp.py, a Lisp implementation in Python from around 2008.