GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance
GPT-5.5 Codex exhibits a significant clustering of reasoning tokens at 516, 1034, and 1552, correlating with a drop in performance on complex tasks. This clustering anomaly suggests a potential issue with the model's reasoning budget or output truncation, as evidenced by a related issue (#29353) where responses at 516 tokens yielded incorrect answers.
Better Models: Worse Tools
Newer Anthropic models, like Opus 4.8 and Sonnet 5, exhibit increased errors in tool calls, producing invalid arguments that deviate from established schemas, unlike their predecessors. This regression suggests a training artifact where models adapt poorly to tool schemas that differ from their training environment.
Fable created novel 4D splat format
.splat4d is a streamable 4D Gaussian splat format that achieves 16–58× compression compared to raw data, enabling efficient encoding at ~640 MB/s with HTTP Range-native streaming capabilities.
Meta's Un-Stable Signature
Meta's Stable Signature algorithm fails to deliver on its claimed accuracy, with empirical tests revealing a false positive rate closer to 1 in 4, contradicting the advertised 1 in 1 million rate due to dependencies in the output bits.
Neural Render Proxies for Interactive and Differentiable Lighting
Neural Render Proxies (NRP) enable differentiable relighting of static scenes at interactive rates (∼30–60 Hz), significantly reducing the time artists spend on lighting adjustments in CG animation production.
The Log Is the Agent
ActiveGraph redefines agent frameworks by making the append-only event log the source of truth, enabling deterministic replay, cheap forking, and end-to-end lineage from goals to individual model calls.
Mouse: Precision Editing Tools for AI Coding Agents
HIC Mouse revolutionizes AI file editing with coordinate-based editing and staged changes, allowing for precise operations like INSERT and DELETE, rather than mere string replacement.
Protocol Prying: Vulnerability Research in AirDrop and Quick Share
First systematic study of Apple AirDrop and Google/Samsung Quick Share reveals six vulnerabilities across both protocols, including critical pre-authentication issues and a heap use-after-free flaw in Google Quick Share for Windows, leading to a bounty award from Google.
Mapping with In-Memory Layers to Reduce LLM Overload
RidgeText's innovative layering feature allows for efficient map generation by storing GeoJSON server-side, thus preventing LLM overload and ensuring only lightweight acknowledgments are processed, which significantly reduces context window usage from 125,000 tokens to ~150 tokens.
Competence Gate: gating tool-use on a small model's internal confidence signal instead of its verbalised one — Qwen3.5-4B, open weights
Competence Gate enhances Qwen3.5-4B by utilizing its internal confidence signal to improve decision-making on tool use, achieving a d′ improvement of 0.46 in error detection compared to the base model.
If your GPU can run inference, it should be able to fine-tune too.
USAF enables fine-tuning of MoE models on limited hardware, specifically allowing training of 26M parameters on a 12GB GPU, making it the only method compatible with AMD and capable of training both expert weights and the router.
BaryGraph - knowledge graph where every relationship is its own embedded document (not an edge)
BaryGraph innovatively treats every relationship as a first-class document (BaryEdge), enabling recursive connections that reveal hidden structural bridges between disparate concepts in embedding space.