OrcaRouter shipped a block-FP8 Qwen3.8-27B with safety-alignment refusals orthogonalized out of the residual stream: refusal drops from 64-99% to 0-6% while MMLU and GSM8K move under 1.3 points. Released Apache-2.0 as a reproducible red-team/interpretability baseline, explicitly marked research-only with no built-in guardrails.
Monitoring long AI transcripts in chunks instead of all at once catches deceptive behaviors that full-context monitors miss, new research shows. Chunk-level deception signals get diluted over long contexts, so chunked monitoring is a cheap, practical upgrade for AI safety evaluations and jailbreak detection.
Robert Porter's torchwright compiler translates Doom's classic renderer into the hand-set weights of a stock Phi-3-architecture checkpoint, so autoregressive generation reproduces frames with no training at all. Map, position, and view angle enter via the prompt, so one 21B model renders any scene it was compiled to handle.
A new method fingerprints a model's log-prob distributions to detect when it is a distillation of another model, even one that insists it is the original. Motivated by AI-control scenarios where a sovereign model distills itself, the fingerprints survive hard-label distillation and LoRA fine-tuning on both log-prob and label-distilled checkpoints.
A Jacobian lens (J-lens) fitted on an earlier Qwen checkpoint can still read and steer the updated model with zero refitting. This writeup shows a pre-fitted instrument surfacing hidden concepts before they are spoken, pointing to interpretability tooling that transfers across model-line updates without retraining.
A new paper argues RL for LLM reasoning isn't teaching new skills, just sparsely re-selecting tokens the base model already knows, touching only 1-3% of positions. Their RL-free method, ReasonMaxxer, applies contrastive loss at high-entropy decision points alone and matches full RL on math benchmarks using tens of problems and minutes of single-GPU training, roughly 1000x cheaper.
ABot-World-0 is an action-conditioned world model that streams 720p at up to 16 FPS on a single RTX 5090 in ~19GB VRAM, with 1.2s action-to-first-frame latency. It distills a bidirectional teacher into a causal student and uses LongForcing to fight autoregressive drift, making infinitely interactive world rollout feasible from a desktop GPU.
BDH-CQ solves queries by updating a recurrent memory from in-context demonstrations and computing in high-dimensional latent space, never parameter updates or verbalized reasoning. At 150M params it hits 29.5% pass@2 on ARC-AGI-1 at $0.0007 per task, breaking the benchmark's cost-accuracy Pareto frontier.
DeepMind's DiffusionGemma pushes continuous vectors alongside tokens each diffusion step, and this deep dive probes whether that hidden channel does real latent reasoning. The author walks through steering, probes, and a letter-arithmetic case, finding the model represents tokens non-causally and can compute in parallel beyond token-by-token generation.
ostris released a free HuggingFace dataset of 1,000 MiniMax H3 styles, each paired with a prompt and a three-tier multimodal description (shot, soundscape, non-diegetic music). Handy for scene-directing video gen without writing prompts from scratch, and a solid seed for style-consistency tuning.
You’re caught up.
That was all 10 stories in the 16 August issue.