Sam's News — tech-research — 2026-09-17¶

AI Safety¶

7.5 Gender and Racial Bias in Open-Weight LLMs for Recruitment¶

A study of six open-weight language models found significant gender and racial bias in recruitment tasks. Agentic job-posting language depressed recruiter recommendations for female candidates, while coded-exclusion language suppressed non-White recruiter scores and deterred non-White personas from applying, operationalizing documented discrimination patterns.

Agentic language reduced female candidate recommendations (r_rb=0.309, p=7×10⁻⁵)
Coded-exclusion language suppressed non-White recruiter scores at large effect sizes (r_rb=0.646–0.758)
Models tested: Llama 3.2, Mistral, Gemma 3, Qwen 3, Phi 3, DeepSeek-R1
Research translates to pre-deployment audit protocol under EU AI Act Annex III high-risk classification and US EEOC adverse-impact analysis
Paper accepted at 9th AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026)

Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS

7 Covert Dialect Bias in Large Language Model Inference¶

A study of ten open-weight language models revealed covert dialect bias despite alignment techniques. Using sociolinguistic methodology, researchers found that AAVE and Nigerian Pidgin were consistently associated with more negative adjectives than Standard American English, with patterns mirroring documented human housing discrimination.

AAVE and Nigerian Pidgin consistently penalized versus Standard American English
Nigerian Pidgin showed most severe penalty across all contexts tested
Tested on 260 meaning-matched sentence quadruples across four English varieties
Bias persists in internal probability distributions despite alignment techniques
Distinct stereotype clusters for each dialect rather than single non-standard category
Paper accepted at 9th AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026)

Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS

6.5 Procedural Traces Shift LLM Oversight Criteria¶

Study shows detailed procedure traces can bias LLM judges toward accepting outputs based on reasoning display rather than accuracy.

Sources: arXiv — Computation and Language RSS

AI Research¶

7 PrimeScientist: Strategic Resource Allocation in Autonomous Research Agents¶

PrimeScientist is a system enabling autonomous research agents to strategically allocate effort across competing scientific directions using adaptive Monte Carlo Tree Search. The approach improved average reward by 10.3% while reducing research attempts by 50.6% compared to baseline methods across 12 AI research tasks.

10.3% improvement in average reward over AutoResearch baseline
50.6% reduction in research attempts under same resource budget
Uses executable plan tree to preserve competing plans and outcomes
Adaptive MCTS-based allocation policy balances exploration and exploitation
Evaluated on AI research, systems optimization, and ML engineering tasks

Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS

6.5 Confidence Estimation from Reasoning to Agents¶

Researchers propose methods for calibrated confidence estimation in language model outputs to improve deployment decisions.

Sources: arXiv — Computation and Language RSS

6.5 SFT vs RL for Tool-Calling Language Model Agents: Controlled Study¶

Researchers conduct a controlled comparison of supervised fine-tuning and reinforcement learning for tool-calling in language model agents.

Sources: arXiv — Computation and Language RSS

6.5 LLM Judge Bias Across Model Families¶

Research finds that choice of model family affects LLM-as-judge evaluation results in a fully crossed design study.

Sources: arXiv — Computation and Language RSS

6.5 Calibrated Measurement Framework for LLM Inference Optimizations¶

Researchers develop a unified instrument to measure how quantization, early-exit, and speculative decoding affect language model output quality.

Sources: arXiv — Computation and Language RSS

6.5 DualSQL: Multi-Agent Reinforcement Learning for Text-to-SQL¶

Method uses multi-agent reinforcement learning to jointly optimize schema linking and SQL generation.

Sources: arXiv — Computation and Language RSS

6.5 Gap Between AI Preference and Real User Engagement Online¶

Analysis of 1.17 million answers reveals misalignment between what language models prefer and what actual users engage with.

Sources: arXiv — Computation and Language RSS