Ant's new guardrail reads the model's own hidden states mid-sentence
A Chinese fintech giant shipped the safety layer most labs only describe in system cards, and a Science Advances paper taught an AI to watch chemical bonds break.
Ant Group has released SingProbe, a guardrail that runs inside the model it polices. Instead of a second classifier re-reading the text, SingProbe reads the base model's own hidden states as tokens are produced, and outputs a ten-dimensional signal covering three jobs at once: what the user is asking for, whether the answer taking shape is unsafe, and whether it is hallucinating. The team describes it in a technical report posted to arXiv and released the training code plus probe checkpoints for several open bases, including its own Ling-3.0 family, under open licences. On offline safety benchmarks the probe averaged an F1 of 0.8561 on the tiny base and 0.8674 on the flash base, beating the strongest standalone guardrail on the paper's own evaluation set; on hallucination detection it reached an average AUC of 0.7765 on tiny, more than three points ahead of the best baseline. The interesting part isn't the score. Running a guardrail inside the decoding loop costs almost nothing extra because the representations are already computed, and Ant argues those same scores can steer generation, not just flag it — its medical variant, SingProbe-Med, intervenes only on the tokens the probe marks as risky, adding roughly 1.8 seconds to a request instead of the 23 seconds a permanently doubled model would cost.
A machine-learning force field has crossed from predicting structures to simulating reactions at scale. MoleculeMind's QuantaMind appears in Science Advances, and the claim is a familiar one in computational chemistry finally being paid off: density-functional-theory-level accuracy at a cost close to classical molecular dynamics. The paper validates reactive trajectories — proton transfer, acid-base neutralisation, phosphate buffering, histidine titration at constant pH — and runs the full catalytic cycle of a PET-degrading enzyme. The authors report stability over tens of nanoseconds and, in post-submission tests, systems past one hundred thousand atoms; the largest configuration actually verified in print is 24,001 atoms over 20 nanoseconds. The catch for anyone hoping to build on it: the code is Apache-2.0, but the weights are non-commercial and released on request. Protein design models have spent two years telling us what a molecule looks like; the useful next question is what it does, which is a motion problem, not a shape one.
HP is shipping Chinese on-device agents preinstalled on its workstations. YuanKong AI, the Peking University spin-out behind the spreadsheet assistant ChatExcel, said it has adapted its 35-billion-parameter on-device model Boxer and its work agent for three HP machines — the ZGX Nano G1n AI workstation, the ZBook 8 G2i mobile workstation and the Z2 Mini G1a desktop — with commercial orders in place since last year and preinstallation the stated direction. Its deployment numbers are the part worth noting: at 40 TOPS of NPU throughput and under 8 GB of memory, the company claims the model retains 87.3% of its task performance. The pitch to enterprises is arithmetic rather than capability — one hardware purchase buys locally generated tokens at near-zero marginal cost, plus data that never leaves the building. HP has no reason to be generous here; a hardware vendor vetting a three-year-old startup for stability, compatibility and multi-year support is a harder filter than any leaderboard.
What to watch: whether an intrinsic probe like SingProbe gets adopted outside Ant's own stack — a guardrail that ships with the base model has distribution no bolt-on classifier can match.
Should a safety guardrail be allowed to edit your answer while the model is still writing it, or only flag it after? Tell us in the comments.
Sources: SingProbe Technical Report (arXiv) · SingProbe (GitHub) · ITBear · Science Advances · QuantaMind (GitHub) · QbitAI · GeekPark · QbitAI
