VeriLoop E2 Operational

VeriLoop E2 is an open-source 27B post-training model from the Robotics Lab at Tsinghua Shenzhen International Graduate School, based on Qwen3.8-27B and designed for code, mathematics, and physics reasoning, with an Apache-2.0 license and native 262K context. VeriLoop E2's core VGR mechanism separates model proposals from external verification and persistence. A candidate state is committed only when protected evidence coordinates do not regress and at least one is strictly improved; otherwise, it is rolled back. The model cannot self-certify success.

Open-source post-training model: A 27B model based on Qwen3.8-27B, released with Apache-2.0 weights and tokenizer, native context of 262,144 tokens, and OpenAI-compatible serving through vLLM 0.17.0.
Code and agent capabilities: Covers repository-level software repair, terminal/tool execution, and other tasks, with reported scores including 76.2% on SWE-bench Pro, 88.8% on Terminal-Bench 2.1, and 64.6% on DeepSWE v1.1.
Mathematical and scientific reasoning: Designed for competition mathematics and challenging scientific question answering, reporting 98.3% on AIME 2026, 100% on ArXivMath, 93.9% on GPQA Diamond, and 89.6% on Apex 2025.
Evidence-closed scientific demonstrations: Releases a rigorous finite-dimensional certificate for a Riemann ζ zero proportion of 67.350003708785593%, along with the T1–T9 framework for asymptotic graviton tomography and an executable certificate suite with 6/6 passing.
Auditable engineering contracts: Publishes 14 Golden Rules that constrain long-running tasks to remain bounded, testable, and traceable, explicitly prohibiting fabricated success, deferred resource budgeting, and overriding local conventions.
Clear boundaries: The production VeriLoop Harness, private orchestration, and training data are not included. Scientific artifacts and the model license are separate, and the results cannot be extrapolated to mean that the Riemann hypothesis or the black hole information paradox has been solved.

Follow WeChat and reply “开源” to join the AI open-source project group
Overall VGR mechanism: VeriLoop-Governed Recurrence models a task-solving process as candidate state transitions around a retained state. The model, acting as a learned proposer, generates candidate latent displacements or artifacts; a fixed decoder materializes them into outputs; and an external verifier returns the protected evidence rank under a fixed contract. An atomic commit occurs only when the candidate rank weakly dominates the incumbent rank and strictly improves at least one coordinate. Otherwise, the incumbent bundle is fully retained, and only the rejected displacement is written to exploration memory.
Protected partial order and atomic commits: The verifier contract defines an ordered obligation list O=(o1​,…,om​) and a rank vector r∈N0m​, where smaller coordinate values indicate fewer remaining violations. The commit predicate χt​=1{rt′​≺P​rt​} requires no coordinate-wise deterioration and at least one strict improvement. Therefore, an improved aggregate score cannot conceal a regression at a single point; incomparable or regressive candidates are always rejected. The commit unit is a CertifiedBundle binding the artifact, latent state, verification result, and summary. When rejected, the entire bundle is retained unchanged, ensuring rollback and replayability. A decreasing integer potential implies that the number of strict commits is finite.
Evidence feedback and rejection exploration: After each attempt, before replacing the incumbent verification object, the evidence compiler Γ calculates bounded, model-visible features from the incumbent result, candidate result, and commit decision, such as rank sums and compressed ranks, improvement/deterioration/parity ratios, and commit and zero-rank indicators. These features are used only to condition the next proposal and have no commit authority. Rejected displacements enter exploration memory Et​ and can influence subsequent search but cannot overwrite the protected state, making failed trajectories “informative but non-persistent.”
Post-training supervision interface: Because VGR labels are defined over task-level state transitions, E2 uses 1,841,831 structured records for post-training. Verified strict-progress actions provide corrective supervision, non-committable candidates provide contrastive negatives, and zero-rank artifacts provide final-generation targets. The loss is a weighted combination of pairwise preference logistic, action NLL, and generation NLL terms, using QLoRA-4/bfloat16, gradient checkpointing, capability-specific routing, and protected-core replay. The report emphasizes that this improves the “proposal distribution” learned by the model.

Obtain the model: Pull tsinghua-sigs-robot-lab/VeriLoop-E2 from Hugging Face, retaining the repository's bundled tokenizer, chat template, and configuration.
Prepare the environment: Install Python 3.12.x, PyTorch 2.10.0 + CUDA 12.9, Transformers 4.57.6, and Triton 3.6.0, then run python -m pip install "vllm==0.17.0".
Start the service: Run VLLM_USE_MODELSCOPE=true vllm serve --served-model-name veriloop-e2 --dtype bfloat16 --model-impl vllm --language-model-only --max-model-len 131072 --max-num-seqs 16 --gpu-memory-utilization 0.92 --generation-config vllm --host 127.0.0.1 --port 8001.
Run a smoke test first: Send a small request with temperature: 0 to http://127.0.0.1:8001/v1/chat/completions and confirm a response such as 8001_OK before routing production traffic.
Call it using the OpenAI protocol: Use /v1/chat/completions with model="veriloop-e2". On the Python side, create a client with OpenAI(base_url="http://127.0.0.1:8001/v1", api_key="EMPTY").
Treat the template as part of the protocol: Do not replace role separators, tool-calling syntax, stop semantics, or reasoning-history formats, as doing so may change tool use and system behavior even when the weights remain unchanged.
Plan for long-context resources: The published tokenizer has a native limit of 262,144 tokens, but the publicly verified service supports 131,072 tokens. Evaluate VRAM and KV cache requirements yourself for longer windows.
Distinguish model and Harness authority: E2 is responsible only for proposals, abstraction, diagnosis, and repair. Evidence acceptance, deterministic checks, commit/rollback, termination, and persistence should be implemented by your own verifier/Harness.

Decoupled generation and verification: The model handles only proposals, diagnosis, and repair, while the external Harness exclusively controls acceptance, rollback, termination, and persistence, preventing the model from self-certifying success.
Regression-resistant commit rules: VGR commits only when protected evidence coordinates do not regress and at least one is strictly improved; an improved aggregate cannot conceal a regression at a single point.
Atomic rollback and safe failure: Rejected candidates retain the complete CertifiedBundle. Failed trajectories can be reused without overwriting verified state.
Supervision from real verification: Post-training constructs preference/NLL objectives from strict progress, non-committable negative examples, and zero-rank final outputs.
Strong general capabilities: At 27B scale, it covers code agents, terminal execution, mathematics, and scientific reasoning, reporting scores including 76.2% on SWE-bench Pro, 98.3% on AIME 2026, and 93.9% on GPQA Diamond.
Long context and open serving: Native 262,144-token context, Apache-2.0 weights, and an OpenAI-compatible serving path verified with vLLM 0.17.0.
Auditable rather than score-only: Evaluation evidence packages bind identity, outputs, native execution/evaluation records, pass/fail results, and integrity metadata to each task.

Hugging Face model repository: https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2
Technical paper: https://openreview.net/pdf?id=P6FIQILHwX

Repository-level code repair: Generate patches for SWE-style tasks, diagnose regressions, and let tests/evaluators decide whether to retain them, making it suitable for maintenance scenarios where modifications must be verifiable.
Terminal and tool agents: Execute commands and call tools in Terminal-Bench-style environments, trigger the next step based on evidence, and use pass/fail credentials to control progression.
Mathematical reasoning and competition solving: Tackle challenging AIME/ArXivMath-style derivations, using external adjudication after multiple rounds of sampling to retain reproducible solutions and reduce the luck of relying on a single answer.
Scientific computing and theoretical physics exploration: Assist with proposing constructions and locating proof gaps in research involving Riemann ζ, quantum field theory, asymptotic gravitons, and related topics.