Tsinghua SIGS Robot Lab Releases VeriLoop E2, a 27B Model with Verifier-Governed State Commitment

Tsinghua SIGS Robot Lab released VeriLoop E2, a 27B parameter model post-trained from Qwen3.8-27B under the Apache-2.0 license. The release centers on VeriLoop-Governed Recurrence (VGR), a design rule for agent state management that prioritizes strict non-regression over scalar optimization.

The VGR Commitment Rule

The VGR rule generates rich training signals by categorizing externally checked candidates into strict progress, no progress, protected regression, incomparable, or zero-rank completion. These categories serve as correction supervision, final-generation targets, or negative examples. The verifier remains outside the gradient path, ensuring the model learns to propose better states without self-authorizing state transitions.

VeriLoop E2 acts as the reasoner/proposer for generation, abstraction, and replanning. The external VeriLoop Harness (not open source) manages execution, comparison, and rollback. The authors clarify that current evidence supports trajectory-supervised realization of VGR, not a full latent-state integration within the Qwen backbone.

Availability and Specifications

• Weights: Apache-2.0, available on Hugging Face.
• Serving: Validated on vLLM 0.17.0 in BF16 at 131K context (tokenizer native max: 262K).
• Quantization: Official GGUF files range from BF16 to IQ1_M. Q6_K (20.6 GiB) is recommended. Note that quant names differ from standard llama.cpp conventions; compare by file size.
• Hardware: The smallest main file is 16.8 GiB, requiring more than a 16 GB VRAM card.
• Evaluation: Per-task records are public, but standalone performance numbers without the proprietary harness are not yet published.