Tsinghua's VeriLoop-E2 splits exploration from commitment, with external evidence gating all state submissions
量子位 · wechat · 2026-10-11
A Tsinghua SIGS team has released VeriLoop-E2, built on 'VeriLoop-Governed Recurrence' (VGR): the model holds exploration rights while a fixed external verification contract holds commitment rights. A candidate is only committed when all protected evidence dimensions are non-regressing and at least one strictly improves (strict Pareto dominance), blocking 'better total score but local regression' updates.
Key points:
- VeriLoopHarness handles task assembly, tool calls, verifier scheduling, rollback and evidence binding; the same evidence semantics organize post-training supervision, teaching the model to propose committable candidates
- 9 public benchmarks covered: repo-level SWE, real terminals, long-horizon code repair, competition math and hard science QA
- Official GGUF quantizations from BF16 to IQ1M released, with PPL/KL fidelity measured against a frozen BF16 reference
- Research demo on the black hole information problem (scoped to asymptotic gravitational radiation reconstruction): 6/6 executable certificates PASS, explicitly not claiming a full paradox resolution
- Developer-facing coding agent 'Voder' is on the roadmap
Weights are on Hugging Face; the technical report is on OpenReview.
More from Research
- Sakana AI proposes MASS: multi-agent self-supervision for recursive self-improvement — hardmaru · 2026-10-13
- Preparing post-training data for coding LLM fine-tunes is 'black art', says practitioner — ayushthakur0 · 2026-10-13
- JQD:DM launches Special Workflow Issue with shared skill-file LLM gatekeeping in peer review — manoelribeiro · 2026-10-13
- Depth vs hold: behavior alone can't diagnose why LLMs fail, new paper argues — burny_tech · 2026-10-13
- BudgetPix: adaptive tokenization matches full-compute pixel diffusion with 25% of the compute — _akhaliq · 2026-10-13
- SoftwareGym synthesizes GUI environments for computer-use agents, lifting OSWorld 37% relative — _akhaliq · 2026-10-13