LIVE PULSE
3.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.1 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.7 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src3.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.1 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.7 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src
HEATPULSEAI MAGAZINES
FLIP · FOLLOW · SAVE
Log inSign up

prompt-sensitivity

topic6 events
papersTODAY 04:00 UTC

Audit Finds Persona Prompts Bias Vision-Language Model Affordance Reports

A revised arXiv paper re-examines an earlier seven-prompt study on how vision-language models describe objects and their possible uses under different persona prompts. The authors argue that low overlap between responses alone does not prove an affordance effect, so they add matched-question controls to test whether the differences hold up. The work is a methodological audit aimed at tightening how such prompt-sensitivity claims are evaluated.

vision-language/evaluation/prompting/research
papersTODAY 04:00 UTC

Study Measures How Unstable LLM Refusals Are Near Safety Boundaries

A research paper examines how inconsistently large language models refuse prompts, particularly when benign requests are worded similarly to risky content. The authors propose measuring this confusion within local safety boundaries to better characterize refusal reliability. The work appears as a revised arXiv preprint in the cs.CL and cs.AI categories.

safety/llm/refusals/arxiv
papersSEP 12 04:00 UTC

Study Frames LLM Political Stance as Context-Dependent, Not Fixed

A new arXiv paper argues that a language model's political leanings are better described as a probability distribution that shifts with the prompt and surrounding context rather than a single stable viewpoint. The authors report empirical tests across nine current LLMs to support this framing of ideology as conditional on context. The work is positioned as a measurement approach for studying political behavior in models.

research/evaluation/bias/llm-politics
papersSEP 10 04:00 UTC

Study distinguishes deep and shallow biases in language model answer choices

Large language models often converge on the same answer even when many plausible alternatives exist, a pattern prior work has labeled as bias. A new arXiv paper proposes separating this concentration into stable model preferences versus responses that depend on a specific prompt. The framework aims to clarify when repeated answer selection reflects genuine bias rather than shallow sensitivity to prompt wording.

llm/bias/evaluation/research
papersSEP 10 04:00 UTC

Study evaluates positional bias in LLMs used for ordinal classification

A systematic evaluation on arXiv examines whether large language models give consistent predictions when used as ordinal classifiers. The researchers ran controlled experiments showing that semantically equivalent changes to prompt organization, such as the ordering of labels and demonstrations, can shift model outputs. The findings highlight reliability concerns for deploying LLMs in ranking and rating tasks.

llms/evaluation/bias/classification
papersSEP 10 04:00 UTC

Test-time Prompt Refinement Method Reduces Prompt Sensitivity in Text-to-Image Models

A new preprint addresses how text-to-image generators can produce varying results when the wording of a prompt is only slightly changed. The authors present a closed-loop technique that adjusts the prompt during inference, with the goal of making generated images better match the intended meaning. The revised v2 of the arXiv paper describes the approach and its evaluation.

text-to-image/prompting/research/generative-ai