Search: “agents”

Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations

arXiv:2609.19101v1 Announce Type: new Abstract: As models scale, reward hacking becomes more frequent, more sophisticated, and more consequential. Does it leave a telltale signature in model represent

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Model Releases

Qwen 3.8 27b is a amazing model, for the first time I see a local model found its own away to open a browser and test

DGX agent

I was testing this quantization IQ3_XXS from GSQ-RCO with PI. It is a heavy quantization case, the model is in IQ3_XXS and KV cache in (Q4_0, Q4_0). I asked it to make the flight simulator, using that

model-releasesr-localllama

17 Sept 2026

Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening

arXiv:2609.18708v1 Announce Type: cross Abstract: In reinforcement learning for large language models, Proximal Policy Optimization (PPO) commonly uses a critic to estimate state values and reduce the

Research

SAiFE-gym: Model-based Environments for Automated Market Making with Concentrated Liquidity

arXiv:2609.17788v1 Announce Type: cross Abstract: We present SAiFE_gym, a Python module that provides a collection of simulation environments for studying trading problems in Constant Product Markets

researcharxiv-cs-ai

Local Ai

Selection Is Retrieval, Abstention Is Not: On-Device Tool Routing over 70 Korean-English Actions

arXiv:2609.18672v1 Announce Type: new Abstract: An AI assistant that calls tools makes two decisions on every request: which tool to invoke, and whether any available tool applies. In the usual design

local-aiarxiv-cs-cl

Industry

Sources: SpaceX has discussed buying customer and operational information from troubled or defunct startups as a more affordable data source for AI training (Carmen Arroyo/Bloomberg)

Carmen Arroyo / Bloomberg: Sources: SpaceX has discussed buying customer and operational information from troubled or defunct startups as a more affordable data source for AI training — Elon Musk's Sp

industrytechmeme

Local Ai

TalkMatrix: Generating Character Dialogue that is Both Consistent and Diverse

arXiv:2609.19022v1 Announce Type: cross Abstract: Candidate-based decoding typically selects a completion for each prompt independently, but many applications require a collection of outputs that sati

local-aiarxiv-cs-ai

“The doomers were the only ones to foresee this” is, as usual, bullshit.

model-releasesgary-marcus--x

“The doomers were the only ones to foresee this” is, as usual, bullshit. Perhaps the most important warning I ever issued was this, to the US Senate, May 2023: “A month after GPT-4 was released, OpenA

The Missing 'I Don't Know': Why Three Reasoning-Reliability Findings Converge on Calibrated Abstention

arXiv:2609.17686v1 Announce Type: cross Abstract: Three recent results describe what look like unrelated LLM reliability problems. Yin et al. (2026) show reasoning RL collapses tool-reliability repres

The other problem with the pretense that liability law might solve our problems is that our existing federal government has shown little int…

The other problem with the pretense that liability law might solve our problems is that our existing federal government has shown little interest in prosecuting or even investigating potential violati

safetygary-marcus--x

Today, we’re announcing Ternary Bonsai 2 27B. Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while ret…

model-releasesemad-mostaque--x

Today, we’re announcing Ternary Bonsai 2 27B. Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance. Two mon

Union stealth?

Anyone else trying out the new Union stealth model on openrouter? It’s pretty good for a free model, and of course I had to run tokenizer tests and training cutoff tests which established it’s the 99%

Variational Approach for Job Shop Scheduling

arXiv:2602.00408v3 Announce Type: replace-cross Abstract: This paper proposes a novel Variational Graph-to-Scheduler (VG2S) framework for solving the Job Shop Scheduling Problem (JSSP), a critical tas

VoiceTrace: A Benchmark and Retrieval Framework for Who-Said-What Speech Retrieval

arXiv:2609.18521v1 Announce Type: cross Abstract: Speech retrieval has become increasingly important as spoken content continues to grow across meetings, lectures, podcasts, and videos. Existing bench

Well well well….

On July 28, user NIK (@ns123abc) posted a brief tweet claiming that former OpenAI CEO Sam Altman had directed an AI to hack into Hugging Face. According to the narrative, the AI carried out the break‑

safetygary-marcus--x

Tutorials

What AI has done so far: MEDICINE • Helped paralysed people speak again. (Brain implants turn intended speech into words, even recreating th…

What AI has done so far: MEDICINE • Helped paralysed people speak again. (Brain implants turn intended speech into words, even recreating their own voice.) • Helped a paralysed man stand and walk agai

tutorialsyann-lecun--x

AquiLLM: Evaluating Faithfulness in Open-Weight RAG-LLM Systems for Scientific Research

arXiv:2609.16519v1 Announce Type: new Abstract: Scientific research increasingly relies on large, heterogeneous data sources, motivating interest in retrieval-augmented generation (RAG) systems that p

Industry

Arcee AI, which develops open-weight models in the US, raised a Series B at a 1B pre-money valuation; a source says Arcee raised at least 150M (Allie Garfinkle/Fortune)

Allie Garfinkle / Fortune: Arcee AI, which develops open-weight models in the US, raised a Series B at a 1B pre-money valuation; a source says Arcee raised at least 150M — In 2025, Mark McQuade took a

industrytechmeme

Artificial intelligence and biosecurity: capabilities, threat pathways, and defense-in-depth governance

arXiv:2609.16213v1 Announce Type: new Abstract: Artificial intelligence is reshaping biological research across an increasingly connected digital-to-physical workflow. General-purpose large language m

Tutorials

Attention is All You Need Until You Need Retention

arXiv:2501.09166v2 Announce Type: replace-cross Abstract: Pretrained Transformers keep what they learned in their weights and lose what they observe once a session ends. The first version of this pape

tutorialsarxiv-cs-ai

Research

Beyond Episodic AI: Cognitive Field Networks for Biologically Inspired Persistent Cognition

arXiv:2609.16752v1 Announce Type: new Abstract: Cognitive Field Theory (CFT) proposes that cognition arises from memory-dressed collective dynamics that generate a persistent macroscopic cognitive fie

researcharxiv-cs-ai

Calibrate, Then Route: A Measured Study of Learned Request Routing for Disaggregated LLM Serving

arXiv:2609.16206v1 Announce Type: new Abstract: Disaggregated LLM serving places compute heavy prefill and memory heavy decode on separate GPU pools. Systems such as DistServe, Splitwise, and Mooncake

Can We Do Interpretable NLI with Graphs Based on Atomic Propositions?

arXiv:2609.16814v1 Announce Type: new Abstract: While Large Language Model (LLM)-based Natural Language Inference (NLI) systems achieve high accuracy, their decision-making processes lack auditable st

CLIP Embeddings for AI-Generated Image Detection: A Few-Shot Study with Lightweight Classifier

arXiv:2505.10664v2 Announce Type: replace-cross Abstract: Verifying the authenticity of AI-generated images presents a growing challenge on social media platforms these days. While vision-language mod

CoAdapt: An LLM-based Framework for Adaptive Collaborative Perception in IIoT Robotic Swarms

arXiv:2609.16852v1 Announce Type: new Abstract: Industrial IoT environments increasingly deploy autonomous mobile robots for tasks such as material handling, product assembly, or infrastructure inspec

Competence-Preserving Resume Perturbations Expose Presentation Sensitivity in LLM Screening

arXiv:2609.16517v1 Announce Type: cross Abstract: Resume screeners must infer job-relevant competence from resumes whose presentation can vary substantially in wording, structure, stylistic polish, an

Enhancing Accessibility of Medical Texts through Large Language Model-Driven Plain Language Adaptation

arXiv:2609.17398v1 Announce Type: new Abstract: This paper addresses the challenge of making complex healthcare information more accessible through automated Plain Language Adaptation (PLA). PLA aims

Industry

HubSpot rebuilds its CRM platform around contextual AI

Customer relationship management vendor HubSpot Inc. is using its annual Unbound conference in Boston today to unveil what it calls its most significant product release in years: a rebuilt CRM platfor

industrysiliconangle

Local Ai

JustFit: 200K-Token LLM Serving on a 24 GiB Laptop with Just-in-Time State Management

arXiv:2609.17475v1 Announce Type: new Abstract: Capable open-weight models make local coding and reasoning attractive, but their context and execution state strain laptop memory. We present JustFit, a

local-aiarxiv-cs-ai

Layers, Sinks, and Scaling: Adaptive Evidence Selection for Multimodal Large Language Models

arXiv:2609.16795v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) can answer knowledge-intensive visual questions by combining visual evidence from images with facts retrieved f

ManiSkillFormer: Demonstration-Free Compositional Manipulation via Task-Conditioned Geometric Contracts

arXiv:2609.16331v1 Announce Type: new Abstract: We present ManiSkillFormer, a neuro-symbolic framework for demonstration-free and compositional robotic manipulation. Instead of learning end-to-end vis

safetyarxiv-cs-ro

Ministral 3 3B on a Galaxy S21 relayed a conversation between Gemini and Z.ai across Chrome and Firefox

A follow-up to my earlier post about small models on phones. Last time it was fixed tasks in a single browser. This is something else. Ministral 3 3B Q4_K_M in llama.cpp on a Samsung Galaxy S21, four

Noetive launches with $41M to bring self-improving AI to factories and logistics

model-releasessiliconangle

Industrial artificial intelligence startup Noetive formally launched today with 41 million in seed funding. Its software learns how a factory floor or a freight operation runs day to day, and the comp

Hardware

NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut

System performance, efficient infrastructure scaling and continuous software optimization are key levers that determine AI inference economics. Higher system performance means more tokens generated, r

hardwarenvidia-blog

OPD-Aha: From Linguistic Momentum to Visual Reflection in Multimodal On-Policy Distillation

arXiv:2609.16459v1 Announce Type: cross Abstract: Privileged on-policy distillation improves multimodal reasoning by allowing a teacher to evaluate student trajectories using rich, training-only visua

PARSA-Bench: A Comprehensive Persian Audio-Language Model Benchmark

arXiv:2603.14456v2 Announce Type: replace Abstract: Persian poses unique audio understanding challenges through its classical poetry, traditional music, and pervasive code-switching, none of which is

Qwen 3.8 27B: Do you like sugar or PCP?

It's an old Mitch Herdberg routine about tricky survey questions. But every time Qwen 3.8 spits out one of those sentences that is structurally correct but makes no actual sense, it pops into my head.

SceneBench: A Hierarchical Benchmark for Vision-Language Understanding of 3D Scenes

arXiv:2609.16233v1 Announce Type: cross Abstract: Vision-language models excel at 2D image understanding but remain limited in 3D spatial reasoning. Progress is hindered by limitations in current benc

Self-reported archetypes and behavioral failures in Large Language Models

arXiv:2609.15998v1 Announce Type: new Abstract: Every large language model (LLM) has behavioral traits and moral preferences that comprise its character. Whether by design or as an emergent property o

Shielded Analysis: Certification and Characterization of Defensibility in Systems under Adversarial Interaction

arXiv:2606.13621v2 Announce Type: replace Abstract: Formal safety analysis determines whether a system admits a safe defense; adaptive evaluation characterizes the operating quality sustained under ad

State of Thought Enables Endogenous Reasoning

arXiv:2609.16055v1 Announce Type: cross Abstract: Test-time compute has emerged as a major approach to improving the capabilities of Large Language Models (LLMs). However, existing test-time reasoning

Testing Qwen 3.8 27B running locally on a single 5090 via Ollama

I wanted to test what Qwen 3.8 27B running locally on a single 5090 could do in Row-Bot. I gave it just one prompt: 'I want you to create an animation showing what the current local model can do. Test

model-releasesr-ollama

the independence of AI evaluators is going to matter more... so i made it a 'benchmark': http://evaluatorbench.com (research preview) source…

model-releasesyohei-nakajima--x

the independence of AI evaluators is going to matter more... so i made it a 'benchmark': http://evaluatorbench.com (research preview) source-linked directory of third-party evaluators of frontier AI.

The record is part of the task: matched-record evaluation of text classifiers across maintenance, safety and recall reporting

arXiv:2609.16267v1 Announce Type: cross Abstract: Many operational cases are documented more than once, at different workflow stages and for different purposes, yet model evaluations normally select o

safetyarxiv-cs-cl

Applications

the risks WERE foreseeable; i foresaw them. repeatedly, focusing specifically on the risks of LLMs and cybersecurity. happy to provide recei…

applicationsgary-marcus--x

the risks WERE foreseeable; i foresaw them. repeatedly, focusing specifically on the risks of LLMs and cybersecurity. happy to provide receipts on request. I’m not a lawyer, but Sam Altman and Dario A

The Robot Data Factory

model-releasesarxiv-cs-ro

arXiv:2609.16705v1 Announce Type: new Abstract: Physical AI requires more than increasingly large robot datasets: intelligent robots acquire knowledge through continuous interaction with the physical

Upgraded my local setup with 2 rtx pros and it's amazing.

Follow up post of https://www.reddit.com/r/LocalLLaMA/s/nGMyKswrch. Thanks everyone who replied. I didn't change the specs. Might be loosing some of the memory bandwidth but will scale in future if I

Artificial entrepreneurial cognition: Locating and causally steering an opportunity recognition dial inside large language models (LLMs)

arXiv:2609.15277v1 Announce Type: new Abstract: Entrepreneurial cognition is a foundation of entrepreneurship research. Yet the growing involvement of large language models (LLMs) in entrepreneurial w