ACL 2026 Publishes Demo Track List, Heavy on LLM Safety Tools

The list of accepted system demonstrations for ACL 2026 is up, and the composition is a decent snapshot of where applied NLP tooling has landed this year. The main conference runs July 2 to 7 in San Diego, so this is what conference-goers will actually be crowding around in the demo hall this week.

The skew is toward LLM safety and agent instrumentation. Titles like Copyright Detective, described on the page as a forensic system for LLM copyright leakage risks, sit alongside Fast-MIA for membership inference, RiskLab for probing emergent risks in multi-agent systems, and DialogGuard for psychosocial safety evaluation of sensitive LLM responses. Interpreto shows up as an explainability library for transformers. On the tooling side there is LiTS for LLM tree search, MixtureKit for composing and visualising mixture-of-experts models, dLLM for diffusion language modeling, and PROTEA for offline evaluation of multi-agent workflows. The domain applications are the other big cluster, with AutoForest generating forest plots from biomedical studies, QFinZero as a financial toolchain for trading agents, Squrve for real-world text-to-SQL, and clinical-facing systems like Dash-M5H and ClinQueryAgent.

Why this matters if you are not going to San Diego: the demo track is where you can usually tell what applied researchers think is worth shipping, not just publishing. The visible pattern here is that safety instrumentation and multi-agent orchestration are no longer novelty topics, they are their own tooling category with enough entrants to fill a room.

The honest caveat is that the ACL page is just a list of accepted titles and authors. It does not publish an acceptance rate, a submission count, or which papers picked up recognition, and it does not tell you which of these tools are already used in production versus one-off research prototypes. Some of these will have durable open source repos, others will not, and the demo format itself is not the place to find that out.

The forward-looking piece is boring but real: for practitioners picking evaluation stacks or shopping for domain-specific NLP collaborators for the second half of 2026, working through this list and pulling the two or three that map onto your workflow is a cheap afternoon.