VLM Errors Trace Back to 'Perception,' Not 'Reasoning'

A research team at UCSC analyzed the causes of vision-language model (VLM) errors and found that **86.9% of Qwen3-VL-8B's wrong answers stem from perception errors**. In other words, the model didn't get things wrong by 'thinking' — it got them wrong by 'seeing.'

Building on this finding, the team proposed a new training method called 'capability curriculum.' The approach trains the model in stages — perception → text reasoning → visual reasoning — and on Qwen3-VL-8B it delivered +1.46% accuracy while shortening reasoning trace length by 20.8%. The key idea is introducing an orthogonal axis — 'which capability to train' — alongside the conventional 'easy → hard' difficulty axis; combining the two reportedly raises the average score from 58.6 to 63.0.

The results held consistently across four backbones: Qwen2.5-VL, Qwen3-VL, InternVL3, and InternVL3.5.

AI Cites the Wrong Passage Even When It Gets the Right Answer

A research team at Peking University flagged a problem where frontier AI models such as GPT and Gemini, when analyzing documents, **produce the correct answer while citing the wrong passage as evidence**. The team named this phenomenon 'attribution hallucination' and released CiteVQA, the first benchmark for systematically evaluating it.

This issue is especially risky in highly regulated fields like law and medicine. Even when an AI reaches the correct conclusion, citing an incorrect source as its basis severely undermines real-world reliability. The team emphasizes the need for mechanisms that verify not just AI outputs at face value, but also the accuracy of their citations.

Running a 1-Trillion-Parameter LLM Locally with 768GB of Optane DIMMs

An enthusiast has shared an attempt to **run Kimi K2.5 (1 trillion parameters) locally on a single GPU** using 768GB of Intel Optane DIMM memory. At roughly 4 tok/s, the speed is far from practical, but the mere fact that such a massive model runs in a local environment is fascinating.

Although Intel Optane DIMMs have been discontinued, they are drawing renewed attention as an option for bypassing the GPU VRAM wall in LLM inference workloads that demand large memory capacities.

MiMo-V2.5-coder: A New Coding Model for 128GB Setups

jedisct1 has released **MiMo-V2.5-coder**, a coding-focused model designed as an alternative to Qwen3.6 and DeepSeek R1 for users with 128GB of memory. It features fast inference and reliable tool calling. A Q2 quantized version is available on Hugging Face.

Cider: W8A8 Quantization Comes to MLX, Speeding Up Apple Silicon

A team at Mininglamp AI developed Cider, an SDK that adds **W8A8 (8-bit weights + 8-bit activations) quantization** to Apple's MLX framework. Previously, MLX only supported weight-only quantization, leaving activations in FP16.

In tests on an M5 Pro (64GB), prefill time for a 4B-parameter VLM dropped from 2.84 seconds to 2.52 seconds. Custom Metal kernels are registered as MLX primitives, which reportedly makes the approach applicable to any model that runs on MLX. Accuracy holds up too: Wikitext2 PPL degradation is minimal (9.71 → 9.76 on Qwen3-8B). Note, however, that compiling INT8 TensorOps requires M5 or later; on M4 and earlier it falls back.

SciAtlas: A Large-Scale Knowledge Graph of Scientific Literature

A large-scale knowledge graph called SciAtlas has been announced, designed to let AI agents go beyond keyword search and understand scientific literature across domains through **structural and topology-based reasoning**. It aims to streamline research understanding that spans multiple academic fields and is expected to serve as foundational infrastructure for AI-assisted academic research.

---