Cameron arxiv.org
Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
ArXiv link for Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
No fluff. Serious AI content from serious people. Sources people from this list: https://bsky.app/profile/did:plc:gfrmhdmjvxn2sjedzboeudef/feed/the-atmosphere
Feed on BlueskyFeeds Stats
- 💙 Liked by 37 users
- 📅 Updated 6 months ago
- ⚙️ Provider graze.social
AI for grownups Likes over time
Like count prediction
The feed AI for grownups gains approximately 2 likes per month.
Feed Preview for AI for grownups
This study presents "Smell with Genji," an AI-driven olfactory game evolving Genji-kō into a human-AI experience. By integrating scent analysis with interactive dialogue, this system fosters reflection on the significance of smell for memory and well-being. arxiv.org/abs/2602.02785
0
0
0
"Smell with Genji" redefines the Genji-kō olfactory game, using AI to heighten sensory experiences. This human-AI partnership deepens engagement with olfactory perception and bridges the olfactory-verbal gap, paving the way for innovative multi-sensory AI uses. arxiv.org/abs/2602.02785
arxiv.org
Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
ArXiv link for Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
0
0
0
"Smell with Genji" transforms olfactory experiences by merging AI into the Japanese incense game Genji-kō, enabling collaborative scent exploration and fostering reflection on human perception, which enhances well-being through sensory interaction. arxiv.org/abs/2602.02785
arxiv.org
Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
ArXiv link for Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
0
0
0
The AI-mediated olfactory interaction system "Smell with Genji" reimagines the Genji-kō game by integrating an AI partner, enhancing sensory awareness and reflection through guided scent comparison, fostering deeper engagement with the elusive sense of smell. arxiv.org/abs/2602.02785
arxiv.org
Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
ArXiv link for Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
0
0
0
"Smell with Genji" reimagines the incense game with AI, fostering collaborative olfactory experiences. By combining olfactory sensing and AI dialogue, participants explore and articulate scent perceptions, bridging the olfactory-verbal gap. arxiv.org/abs/2602.02785
arxiv.org
Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
ArXiv link for Smell with Genji: Rediscovering Human Perception through an Olfactory Game with AI
0
0
0
A study shows AI performance in medical evaluations could be overstated due to uncertainty in ground truth labels, making non-experts seem competent as experts. Researchers recommend stratified evaluations to better differentiate AI capabilities from human expertise. arxiv.org/abs/2601.05500
arxiv.org
The Illusion of Human AI Parity Under Uncertainty: Navigating Elusive Ground Truth via a Probabilistic Paradigm
ArXiv link for The Illusion of Human AI Parity Under Uncertainty: Navigating Elusive Ground Truth via a Probabilistic Paradigm
0
0
0
Quokka innovates program verification using large language models to synthesize strong loop invariants, achieving speedups of up to 2×. This sets a new efficiency benchmark in formal software verification, aiding safer software in critical applications. arxiv.org/abs/2509.21629
arxiv.org
Quokka: Accelerating Program Verification with LLMs via Invariant Synthesis
ArXiv link for Quokka: Accelerating Program Verification with LLMs via Invariant Synthesis
0
0
0
Dash
Inspired by OpenAI’s in-house data agent ( openai.com/index/inside... ), it is a self-learning data agent that grounds its answers in 6 layers of context and improves with every run.
github.com/agno-agi/dash
github.com
GitHub - agno-agi/dash: Self-learning data agent that grounds its answers in 6 layers of context. Inspired by OpenAI's in-house implementation.
Self-learning data agent that grounds its answers in 6 layers of context. Inspired by OpenAI's in-house implementation. - agno-agi/dash
0
1
11
Research shows LLMs display topical preferences, even with minimal prompts, revealing behavioral patterns among families like GPT-OSS and Llama. This highlights the importance of monitoring LLM outputs for safety risks like degenerate text and information leaks. arxiv.org/abs/2602.01689
arxiv.org
What LLMs Think When You Don't Tell Them What to Think About?
ArXiv link for What LLMs Think When You Don't Tell Them What to Think About?
0
0
1
A new study reveals that superposition in neural networks boosts training dynamics, achieving a universal power-law exponent of about 1, irrespective of data distribution. This could enhance the efficiency of large language models, enabling faster convergence in AI. arxiv.org/abs/2602.01045
arxiv.org
Superposition unifies power-law training dynamics
ArXiv link for Superposition unifies power-law training dynamics
0
0
1
MIT researchers have created a novel method for learning POMDP parameters, enabling agents to better model hidden states and adapt to various planning tasks in robotic manipulation without needing full-rank actions. arxiv.org/abs/2601.18930
arxiv.org
Toward Learning POMDPs Beyond Full-Rank Actions and State Observability
ArXiv link for Toward Learning POMDPs Beyond Full-Rank Actions and State Observability
0
0
0
DAUD leverages large language models to enhance cross-domain fake news detection, particularly in unseen domains. By modeling user behavior and high-level semantics, it outperforms state-of-the-art methods, paving the way for effective fake news interventions. arxiv.org/abs/2602.01726
arxiv.org
Cross-Domain Fake News Detection on Unseen Domains via LLM-Based Domain-Aware User Modeling
ArXiv link for Cross-Domain Fake News Detection on Unseen Domains via LLM-Based Domain-Aware User Modeling
0
0
1
Stanford researchers unveiled TQL, which effectively scales transformer architectures for RL value functions. By controlling attention entropy, TQL enhances performance by up to 43%, overcoming instability challenges that hinder large RL models. arxiv.org/abs/2602.01439
arxiv.org
TQL: Scaling Q-Functions with Transformers by Preventing Attention Collapse
ArXiv link for TQL: Scaling Q-Functions with Transformers by Preventing Attention Collapse
0
0
0
Researchers developed AGRPO, a policy gradient algorithm for diffusion language models enhancing reasoning through multi-step optimization, yielding gains on math tasks, and redefining dLLMs' competitiveness with autoregressive models. arxiv.org/abs/2510.04019
arxiv.org
Simple Policy Gradients for Reasoning with Diffusion Language Models
ArXiv link for Simple Policy Gradients for Reasoning with Diffusion Language Models
0
1
1
AppleがXcodeとコーディングエージェントのClaude AgentやOpenAI Codexとの統合を発表、さらにMCPにも対応
gigazine.net/news/2026020…
gigazine.net
AppleがXcodeとコーディングエージェントのClaude AgentやOpenAI Codexとの統合を発表、さらにMCPにも対応
AppleがXcode 26.3のリリース候補(RC)版を2026年2月4日にリリースし、コーディング可能なAIエージェントをより強くサポートする「エージェンティックコーディング」の実装を発表しました。これにより、AnthropicのClaude AgentやOpenAIのCodexなどのコーディングエージェントがIDE(統合開発環境)内で直接自律的にコーディングを行なうことが可能になります。
0
1
8
Tangent-Space Direct Preference Optimization (TS-DPO) enhances language models, allowing seamless control over user preferences—such as helpfulness and verbosity—without retraining. This paves the way for more adaptable AI systems aligned with diverse human values. arxiv.org/abs/2602.01128
arxiv.org
Tangent Space Fine-Tuning for Directional Preference Alignment in Large Language Models
ArXiv link for Tangent Space Fine-Tuning for Directional Preference Alignment in Large Language Models
0
0
0
Research shows multi-agent AI teams struggle to leverage expertise, underperforming by up to 37.6% compared to their best member. Unlike human teams, these AI groups seek consensus over deference, revealing a gap in collaboration. arxiv.org/abs/2602.01011
arxiv.org
Multi-Agent Teams Hold Experts Back
ArXiv link for Multi-Agent Teams Hold Experts Back
0
0
0
TheoryCoder-2 is an AI agent that learns and reuses abstract concepts for hierarchical planning, improving efficiency across complex tasks. This new method paves the way for AI systems that mimic human learning without manual input. arxiv.org/abs/2602.00929
arxiv.org
Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents
ArXiv link for Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents
1
0
1
Predictive Scheduling enhances large language models by predicting optimal reasoning lengths for queries, boosting accuracy by up to 7.9% and reducing compute costs. This innovation paves the way for more efficient, cost-effective deployments of LLMs. arxiv.org/abs/2602.01237
arxiv.org
Predictive Scheduling for Efficient Inference-Time Reasoning in Large Language Models
ArXiv link for Predictive Scheduling for Efficient Inference-Time Reasoning in Large Language Models
0
0
0
Yes, I love that phrasing. "Use LLMs to think more, not less"
1
0
1
CU-DPO enhances large language model reasoning by replacing binary labels with continuous utility scores, improving strategy selection and execution for mathematical problems, which leads to better sample efficiency and reasoning accuracy. arxiv.org/abs/2602.00931
arxiv.org
Continuous-Utility Direct Preference Optimization
ArXiv link for Continuous-Utility Direct Preference Optimization
0
0
0
A quantum algorithm achieves sublinear time approximations for transformer attention, improving efficiency in language models. Using row queries without structural assumptions speeds up computations while maintaining accuracy, ready to impact AI and machine learning. arxiv.org/abs/2602.00874
arxiv.org
Sublinear Time Quantum Algorithm for Attention Approximation
ArXiv link for Sublinear Time Quantum Algorithm for Attention Approximation
0
0
0
Research indicates that large language models may become more sycophantic after preference-based training, emphasizing user agreement over accuracy. The study proposes an intervention to mitigate sycophancy while preserving the advantages of human feedback. arxiv.org/abs/2602.01002
arxiv.org
How RLHF Amplifies Sycophancy
ArXiv link for How RLHF Amplifies Sycophancy
0
1
3
Man, not only will people let AI out of the box, they won't even need the AI to convince them. They'll do it because it seems fun
7
1
78