Watermarking code by only coloring grammar-legal tokens, then turning up their temperature, because syntax was getting in the way.
trending16
01 02 Rethinking Visual Provenance: Detection and Watermarking Across Direct Visual Generation and LLM-Driven Code Rendering arxiv.orgHow do you prove an image came from a model, whether it was generated directly or rendered from code?03 arXiv paper on Reed-Muller LM watermarking that survives deletions, because token alignment was the real enemy.04 Semantic Behavioral Watermarking: Paraphrase-Robust and Forgery-Resistant Provenance for LLM Agents arxiv.orgAction-level watermarking for LLM agents, meant to survive paraphrases and make forgery need the key.05 Watermarking is declared impossible against adaptive removal, so the paper pivots to auditable compliance via robustness profiles.06 Multi-bit text watermarking with f-divergence budgets, via coding-based next-token biasing and the usual rate vs detectability trade.07 Watermarking can make LLMs forget facts they were handed, and the paper tests six schemes plus fixes.08 SEW watermarks LLM code by hiding a secret style dialect in syntax, then tests it against edits and false positives.09 TTMark swaps single-token watermarking for adjacent-token pairs, trying to squeeze signal out of low-entropy text.10 AutoMark uses frontier models to autosearch LLM watermarks, finding 50+ schemes and a few that beat the usual tradeoffs.11 Beyond Semantic Narrowing: Robust and Efficient LLM Watermarking with Hamming Neighborhoods arxiv.orgArXiv paper on HammingMark, a semantic LLM watermark that swaps semantic narrowing for a Hamming neighborhood.12 Paper models image watermarking as AWGN/BPSK, then notes current schemes are keyless and weak, and proposes SNW.13 ArXiv paper on LoRo-Mark, a black-box watermark for repackaged LLM agents, because ownership disputes weren't messy enough.14 Can speculative sampling stay exact and watermarkable? This paper says yes, via Poisson processes and drafter-invariant list coupling.15 A survey of pseudorandom codes, plus the new generic strengthening tricks and open problems the field apparently needed.16 Dual-Locking Learned AI Models: A PIN-Based Sparse QIM Watermarking and Adaptive Index Permutation Approach arxiv.orgTwo-lock watermarking for neural nets, sparse QIM plus key permutation, so ownership checks and recovery need the right PIN.