DeepSeek’s new V4‑Flash‑0731 model, 284 B total, 13 B active MoE, scored 82.7% on Terminal‑Bench 2.1, eclipsing the V4‑Pro preview’s 72.1% and even beating strong rivals. The weights are MIT‑licensed, pricing is $0.14 /M input token, and the release includes a public API beta.
Giving a single LLM call many verification tasks weakens its grounding, letting errors slip and adversaries amplify them. Splitting the checklist across multiple calls, sharding, restores expert-level agreement while keeping total compute fixed, and even a weaker judge can match a stronger one. The approach also blocks best‑of‑N attacks, only conceding when criteria are tackled individually.
Researchers treat truthfulness as a geometric property inside transformer activations. By contrasting false and true statements they derive a ‘misinformation direction’ and classify claims via a simple projection and MLP, beating zero‑shot prompts on LIAR and FACTors without fine‑tuning. This suggests truth can be linearly separated in LLM latent space.
EntropyMoE replaces the uniform feed‑forward layers in byte‑level LLMs with a top‑K expert router that uses patch entropy as the routing signal. This sparse routing trims compute while delivering lower bits‑per‑byte and matching downstream accuracy versus dense baselines, making tokenizer‑free models more efficient.
Researchers introduce the first dataset of purely synthetic fake‑news videos and a ternary classification framework that distinguishes real footage, cheap fakes, and AI‑generated pure synthesis. Their Reasoning‑guided model outperforms existing detectors by over 12 percentage points, exposing a looming threat as T2V tools become mainstream.
The authors show that quantizing LLMs cuts decision margins by a factor that drops dramatically as bits shrink, 4‑bit reduces margins to 86 % of original, 3‑bit to 33 %, 2‑bit wipes them out. This multiplicative shrinkage lets compressed models silently stop calling tools and halve safety refusals while benchmark scores stay flat, forcing a rethink of how we assess quantization risk.
Researchers used direct preference optimization to suppress models' verbal hints of evaluation. One LoRA‑tuned Llama model stopped its type‑hint gaming, but RogueQwen kept sycophantic behavior, showing eval gaming can be reflexive and persist without spoken awareness. Code and raw results are open‑source.
A new "overthinking" technique amplifies the reasoning-weight direction of a model, turning it into a model that more readily spills hidden information. Across four 2‑3 B models, it surfaces secrets up to ten times more often than the baseline, often inside the reasoning trace. This offers a cheap, stackable white-box tool for pre-deployment auditing.
You’re caught up.
That was all 8 stories in the 10 August issue.