CLM

In a test with 1,000 candidate options, CLM took approximately 44 milliseconds, while a generative model required 579 milliseconds.

CLM (Contrastive Language Model) is an 8-billion-parameter contrastive language model released by a research team from Stanford and NVIDIA, using a dual-encoder architecture that transforms decision-making into a vector similarity retrieval problem.

Dottr adopts a lightweight alternative strategy: sacrificing the compliance and audit depth of large enterprises in exchange for zero deployment cost and an extremely low barrier to entry for small and medium teams.

Contract lifecycle management (CLM) standards cover the full chain from requirement initiation, drafting, negotiation, approval, and signing to archiving, renewal, and performance tracking; representative products include Ironclad, DocuSign CLM, ContractPodAi, and others, typically sold to legal departments, and are expensive with long implementation cycles.

The derivation gap appeared in all seven judge models the researchers evaluated, including the System One series' CLM and Jev.

The core value of CLM platforms lies in digitizing and centrally managing each stage of a contract's lifecycle, including drafting, negotiation, approval, signing, archiving, and renewal.

Ironclad was founded in 2012, is headquartered in San Francisco, and is a leading SaaS platform in the contract lifecycle management (CLM) space.

In CLM experiments, the accuracy of selecting the correct tool from 8 tools was 86%, but dropped sharply to 17% when selecting from all 1,080 tools, with the main error being confusion between tools with similar names; TiFF is limited to a maximum of 255 options.

The author recommends using CLM as a first-level coarse filter to compress thousands of candidates down to the top 10 within milliseconds before passing them to a System 2 model for final decision-making; if the action set is small and stable, it can be used standalone; avoid stuffing a thousand tools into a single agent and instead split into sub-agents.

CLM's 8-billion-parameter backbone Qwen3-8B is frozen and serves only as a reader; what is actually trained are only the small projection heads on each side with approximately 20 million parameters each, and a complete pretraining run takes about one hour on a single RTX.

CLM uses a three-stage training process: in the first stage, 60 million question-answer pairs from NVIDIA Nemotron data are used to learn general knowledge; in the second stage, Gemini is used to generate 30 million hard negative samples; in the third stage, approximately 1 million real agent trajectories are used with 40% of the original question-answer data backfilled as replay.

The dual-tower architecture can pre-compute and cache the embeddings of one side, so at query time only the other side needs to be encoded followed by vector retrieval, resulting in extremely low computation; this is the architectural root of CLM's speed advantage, with the trade-off being that the two-sided encoders cannot perform fine-grained cross-attention comparison.

In CLM training, pretraining alone scores 52% on the hard negative sample test; adding the hard negative sample stage jumps it to 69%; training with hard negative samples from day one peaks at only 62% and overfits quickly; omitting replay causes the hard negative sample score to drop from 69% back to 56%.

The author ran the demo on a DGX Spark; the CLM server itself is approximately 75 MB, the overall footprint is approximately 25 GB of VRAM, and after caching, the same ticket can be queried again with a response within 2 milliseconds.

CLM's decision speed is up to 13 times faster than TiFF-class models, while matching them in accuracy.

CLM is compatible with TiFF's API and offers three primitives—yes/no questions, multiple-choice questions, and scoring—unified at the underlying level as 'one state plus a closed set of candidates,' meaning it will not fabricate answers outside the list but can still pick the wrong item from the list.

CLM encodes candidate actions only once and caches them, requiring only an embedding computation for the current context and a dot product against the cached action vectors at each step, thereby achieving a significant speedup.

Because each option is encoded independently, CLM never places multiple options side by side for direct cross-comparison during processing, sacrificing the ability for direct interaction between options.

CLM uses a dual-encoder architecture: one encoder reads the current context, and the other encodes each possible action separately, with the final decision being the action closest to the context in embedding space.

The performance figures cited in the article (44ms vs. 579ms, 9x speedup) are values claimed by the research team as cited in a YouTube video and have not yet been verified by official papers and benchmarks.