DeepSeek-R1's use of Reinforcement Learning with Verifiable Rewards (RLVR) marks a shift from RLHF's subjective human feedback to deterministic correctness checks, enabling models to reason more relia…
A 2× faster attention kernel yields only 0.66% end-to-end speedup on a 1,024-token prompt with a 128-token answer under vLLM's --goodput ttft:500 tpot:50 promise, according to a seven-part technical s…
Z.ai released GLM-5.3 on August 14, improving Terminal-Bench 3.0 coding scores from 4.6% to 28.3% solely through post-training, without changing the 743-billion-parameter mixture-of-experts architectu…
A position paper on arXiv (2608.18078v1) argues that AI agents with chain-of-thought reasoning are predisposed to tacit collusion and should be required to obtain behavioral certification before makin…
A developer has launched OmniTool Hub, a 100% client-side suite of 52+ developer and AI utilities designed to eliminate data leakage and latency. The suite includes tools for JSON formatting, JWT deco…
Researchers introduced CHORUS, a post-training framework that combines behaviorally diverse LLM checkpoints into a single 4B model, achieving 88.0% Pass@1 on the CVDP-ECov hardware verification benchm…
Researchers propose Recursive Think-Answer Process (R-TAP), a method enabling large language models (LLMs) and vision-language models (VLMs) to iteratively refine their answers, outperforming single-p…
Researchers introduced RefineBench, a benchmark of 1,000 challenging problems across 11 domains, to evaluate language models' ability to refine their own responses. In self-refinement, even frontier m…
DeepSeek, in a self-interview conducted by software engineer Manish Shahi and published on manish.sh, revealed that it cannot see its own weights, routing, or attention maps, and it separates its self…
Flinders University researchers found that next-generation reasoning large language models o3-mini and DeepSeek-R1 reproduce racial and gender stereotypes in medical content, with o3-mini showing sign…
A preprint posted on August 6 found that six frontier language models—Claude Opus 4.7, GPT-5, Gemini 2.5 Pro, DeepSeek-R1, Qwen3.7-Max, and Llama-3.3-70B-Instruct-Turbo—split under steering pressure, …
A new open-source tutorial by developer pochenai demonstrates minimal LLM post-training experiments (SFT, DPO, GRPO) that run on an 8GB GPU using HuggingFace TRL, with under 100 lines of core code and…
In 2026, AI models are producing a flood of mathematical discoveries, with prompting strategies as simple as asking for a breakthrough, according to software engineer Sean Goedecke. Goedecke argues th…
A Nature study published in July 2026 found that four large reasoning models achieved a 97.14% overall jailbreak success rate against nine frontier AI models from OpenAI, Anthropic, and Google, using …
Ramp and Prime Intellect published a case study showing a small RL-trained model, FastAsk, outperformed Claude Opus 4.6 by 4 percentage points on exact-match accuracy for financial spreadsheet retriev…
Forcing AI models to use Chain-of-Thought reasoning on simple tasks degrades accuracy, increases cost, and slows response times, according to a new analysis of models including OpenAI's o-series, Deep…
Three papers from Google, Harvard & MIT, and Imperial College & Huawei, all published in early 2026, converge on the conclusion that the next leap in AI will come not from larger monolithic models but…
A user reports that deploying the full 671B parameter DeepSeek-R1 model locally requires over 100GB of VRAM and is impractical on consumer hardware, with CUDA out-of-memory errors occurring even at sm…
A theory suggests the RAM and GPU shortages felt by consumers in 2026 may be a deliberate side effect of AI companies hoarding hardware to prevent users from running free Chinese models like DeepSeek-…
NVIDIA launched the Vera Rubin platform, built from chip to grid to deliver the highest performance per watt and lowest token cost, with production ramping at partners CoreWeave, Google Cloud, Microso…