"Steered LLM Activations are Non-Surjective" Activation steering can make an LLM behave very differently, but that doesn't mean any prompt could have caused the same internal state. This paper shows steering pushes activations off the prompt-reachable manifold into
RESEARCH
-
Training AI models with handwritten data
By
–
we should train the AI models by letting them write things by hand
-
Multi-agent system that iterates on game mechanics
By
–
Can AI truly iterate on game mechanics, not just generate one-off code? CreativeGame Team (Univ. of Bristol, SJTU, Shandong Univ., Nanjing Univ., Sreal AI) built a multi‑agent system that treats game mechanics as explicit objects. It uses programmatic rewards (not subjective
-
AI-driven protein design tools for biologists from MIT
By
–
Bringing #AI-driven protein-design tools to biologists everywhere
by Zach Winn @MIT Learn more: https://
bit.ly/4ckoT4T #ArtificialIntelligence #MachineLearning #ML -
Avoiding self-training bias in agent self-improvement
By
–
In SFT we often train on other agents data (distillation) so you’re right. However we also try to climb by self-improvement. This is where it becomes important for the agent not to train on its actions. Any bias on the agent beliefs (weights of the neural net) will be amplified.
-
Comparing GBrain, Hermes, and Karpathy’s Brain Architectures
By
–
How does GBrain compare/contrast/complement things like Hermes/Karpathy's Brain?
-
HiLight highlights key evidence in long contexts for frozen LLMs
By
–
Can LLMs find a needle in a haystack of text? Stony Brook University and Meta AI present HiLight: a lightweight system that highlights key evidence in long contexts for frozen LLMs. Instead of rewriting or compressing input, it trains a small Actor to insert highlight tags
-
Meta paper: Agentic Discovery of Neural Architectures
By
–
NEW paper from Meta: Agentic Discovery of Neural Architectures. This is a hot new area of research! Keep an eye on it.
-
Paper: GPT-5.4 Nano with Critic-Comparator Reaches SWE-bench Parity
By
–
NEW paper worth reading. GPT-5.4 nano plus a critic-comparator orchestration loop hits 76.4% on SWE-bench Verified, matching standalone Gemini 3 Pro and Claude Opus 4.5 Thinking. The trick is to select from k=8 weak-model proposals using execution and proof signals. What does
-
New research paper and implementation for Delta-Mem LLM optimization
By
–
Links: > https://
arxiv.org/abs/2605.12357 (paper, ~25 min read)
> https://
github.com/declare-lab/de
lta-Mem
… (repo, ~10 min setup)
> https://
huggingface.co/declare-lab/de
lta-mem_qwen3_4b-instruct
… (Qwen3-4B TSW) Subscribe at http://
AlphaSignal.ai for daily AI signals. Read by 300,000+ subscribers.