Proving what is real: watermarks that survive paraphrase and detectors that run on chip
Four new systems attack the same problem from opposite ends: mark model output so it can be checked later, and detect synthetic media fast enough to run everywhere.
- KVMark detects KV-cache tampering with 100% detection and a 0% false positive rate, with 100% layer-level localization. Self-reported m:…1449
- SAMark holds 90.2% detection at a 1% false-positive rate under paragraph-level paraphrasing attacks. Self-reported m:…1445
- WeVisForensics reaches 95.2% detection accuracy for AI-generated video across in-domain, out-of-domain, and agentic benchmarks. Self-reported m:…7415
- Lasso reports an 83,784x speedup and over 99% energy reduction against software-only detectors. Self-reported m:…9776
Mark it where it lives
The first system marks the model's own working memory. KVMark is a training-free, zero-bit watermarking framework for KV-cache tensors that enables secure, verifiable communication between agents in latent multi-agent systems. The paper reports 100% detection and a 0% false positive rate for KV-cache tampering in multi-agent pipelines, with 100% layer-level tamper localization. It maintains reasoning accuracy across benchmarks without requiring model re-training. m:…1449
The idea is to watermark the channel, not just the output. When agents pass each other half-finished reasoning in latent form, a tampered tensor is invisible to every check that only reads the final text. Marking the cache itself makes the handoff checkable. m:…1449
The multi-agent setting is what makes the KV-cache result matter beyond a single model. KVMark is built for latent multi-agent systems, where agents hand each other internal state rather than finished text, and its claim is that this handoff can be made secure and verifiable. A pipeline in which one agent's half-finished reasoning becomes another agent's input needs exactly this kind of check: without it, a tampered tensor propagates silently through every agent downstream. m:…1449
Survive the paraphrase
The second system marks text so it survives rewriting. SAMark, a self-anchored text watermarking framework, achieves 90.2% detection at a 1% false-positive rate under paragraph-level paraphrasing attacks, and outperforms the strongest prior semantic-level watermarking baseline by more than 30% on average. It uses a step-independent green region in semantic space to maintain watermark integrity despite sentence reordering. m:…1445
Detection gets an agentic upgrade in a third paper. WeVisForensics, a self-evolving agent for detecting AI-generated video, achieves 95.2% detection accuracy across in-domain, out-of-domain, and agentic benchmarks. Its dual-timescale policy optimization improves forensic capability by closing the loop between investigator and inspector agents, and its ForenCue-30K dataset provides annotated forensic intervals and interaction trajectories for training. m:…7415
The video detector's training data is part of its claim. ForenCue-30K provides annotated forensic intervals and interaction trajectories for training forensic agents, which is what lets the investigator and inspector agents learn from each other's work rather than from a static set of labeled clips. The dual-timescale optimization closes the loop between the two roles, so the system improves at finding the traces that matter instead of memorizing the artifacts of one generator. m:…7415
Fast enough to run everywhere
The fourth system attacks cost. Lasso, a hardware-accelerated detection method using Intel QAT on-chip compression, achieves 100% accuracy on original DetectRL text and 92% after paraphrasing. The paper reports an 83,784x speedup and more than 99% energy reduction compared with software-only detectors. The same work notes that the Intel QAT microarchitecture exhibits data structure-dependent timing properties that can be used as a measuring instrument for AI detection. m:…9776
The four systems do not compete; they cover different checkpoints. One marks the agent-to-agent channel, one marks the text, one inspects the video, and one makes detection cheap enough to run at the edge. A provenance story that relies on any single one of them is thinner than one that layers them. m:…1449 m:…9776
The speed number is the one that changes who can run detection at all. A detector that is 83,784x faster and uses more than 99% less energy than software-only detectors stops being a forensic lab tool and becomes something that can sit in the path of ordinary traffic. That is the point of building it on Intel QAT on-chip compression: detection that only works in a data center cannot catch synthetic media where people actually encounter it. m:…9776
What is still unresolved
- Whether KV-cache watermarking survives the compression and quantization steps of real serving stacks
- Whether paraphrase-robust watermarks hold against human rewriting, not just model paraphrase
- How detection accuracy degrades on video from generators newer than the training data
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.