Releases·Model Release
Tsinghua SIGS lab releases VeriLoop E2, a 27B Apache-2.0 Qwen3.8 post-train with verifier-gated state commits
A Tsinghua SIGS robot lab team released VeriLoop E2, a 27B model post-trained from Qwen3.8-27B under Apache-2.0, alongside official GGUF quants and a technical report. The release centers on VeriLoop-Governed Recurrence (VGR), a commit rule that only lets the model replace verified agent state when the new state strictly improves a protected per-obligation ranking. The external verifier harness that enforces the rule and produced the benchmark numbers is not open-sourced.
22 engagement1 sourceSun, Sep 27, 2026, 03:03 PM
AI Overview
- VeriLoop E2 is a 27B model post-trained from Qwen3.8-27B, released Apache-2.0 with weights, tokenizer, config, and vLLM inference utilities
- The VeriLoop-Governed Recurrence commit rule only accepts proposed states that worsen no protected obligation and improve at least one, giving the model proposal authority but not persistence authority
- The same commit rule yields training labels — strict progress, no progress, protected regression, incomparable, zero-rank — with the verifier kept outside the gradient path
- E2 evidence covers only trajectory-supervised VGR; the authors say a full latent-state realization in the Qwen backbone lacks supporting evidence
- Official GGUFs span BF16 to IQ1_M; Q6_K (20.6 GiB) is recommended, and quant tiers below Q6_K keep higher precision than standard llama.cpp names
- Serving validated on vLLM 0.17.0 in BF16 at 131K context; tokenizer maxes at 262K; the production VeriLoop Harness stays closed source
Related
Social·Community
Tsinghua SIGS lab releases VeriLoop E2, a Qwen3.8-27B Apache-2.0 post-train, with no official GGUF yet
A Reddit post highlights VeriLoop E2, a newly released Apache-2.0 27B model post-trained from Qwen3.8-27B for code, math, scientific reasoning, and agentic workloads. The release ships full safetensors weights on Hugging Face but no official GGUF/Ollama build, prompting discussion of which quantization and deployment targets the local-inference community should prioritize.
A Reddit post highlights VeriLoop E2, a newly released Apache-2.0 27B model post-trained from Qwen3.8-27B for code, math, scientific reasoning, and agentic workloads. The release ships full safetensors weights on Hugging Face but no official GGUF/Ollama build, prompting discussion of which quantization and deployment targets the local-inference community should prioritize.
6 engagement1 sourceredditWed, Sep 23, 2026, 12:43 PM
Social·Community
VeriLoop E2 vs Qwen 3.8 27B Base, Another Benchmaxxed slop or?
A user tested VeriLoop E2, a 27B finetune, against Qwen 3.8 27B variants on a website generation task. Their 3-bit GSQ-RCO quant of VeriLoop E2 produced the best result, using the fewest tokens and running fastest. The user released the quant on Hugging Face.
A user tested VeriLoop E2, a 27B finetune, against Qwen 3.8 27B variants on a website generation task. Their 3-bit GSQ-RCO quant of VeriLoop E2 produced the best result, using the fewest tokens and running fastest. The user released the quant on Hugging Face.
3 engagement1 sourceredditMon, Sep 28, 2026, 11:16 AM
Developments·Paper
Verification as an Architectural Layer for LLM Agents: A V-Model Design, and a Pilot Study of Its Deterministic Core
An arXiv paper adapts the software-engineering V-model to LLM agents, treating verification as an architectural layer with specification levels, dedicated verifier pairs, and a deterministic controller that enforces every verdict. In a pilot on MuSiQue's four-hop stratum with an 8B backbone, unverified configurations answered zero of ten questions while the verified no-planner configuration answered eight and abstained on the rest. The work targets a core agent failure mode: ReAct agents that cannot progress run until an external budget stops them instead of reporting failure.
An arXiv paper adapts the software-engineering V-model to LLM agents, treating verification as an architectural layer with specification levels, dedicated verifier pairs, and a deterministic controller that enforces every verdict. In a pilot on MuSiQue's four-hop stratum with an 8B backbone, unverified configurations answered zero of ten questions while the verified no-planner configuration answered eight and abstained on the rest. The work targets a core agent failure mode: ReAct agents that cannot progress run until an external budget stops them instead of reporting failure.
0 engagement1 sourcearxivFri, Sep 25, 2026, 07:30 PM
Releases·Tool Release
Two new quantization releases target Qwen3.8-27B: GSQ-RCO GGUFs and QUASAR NVFP4
Two independent teams released aggressively quantized versions of Qwen3.8-27B within two days. One ships GGUF checkpoints at 2.5-3.0 bits per weight using two new methods, GSQ and RCO, with both the search and quantizer learned. The other releases an NVFP4 W4A4 checkpoint trained via quantization-aware distillation with the QUASAR QAT algorithm, deployable on vLLM with NVIDIA Blackwell GPUs. Both matter to practitioners running large models at low precision locally or on modern accelerators.
Two independent teams released aggressively quantized versions of Qwen3.8-27B within two days. One ships GGUF checkpoints at 2.5-3.0 bits per weight using two new methods, GSQ and RCO, with both the search and quantizer learned. The other releases an NVFP4 W4A4 checkpoint trained via quantization-aware distillation with the QUASAR QAT algorithm, deployable on vLLM with NVIDIA Blackwell GPUs. Both matter to practitioners running large models at low precision locally or on modern accelerators.
Updated Fri, Aug 28, 2026, 09:46 PM — GSQ-RCO GGUF release at 2.5-3.0 bpw followed the earlier QUASAR NVFP4 checkpoint on 2026-08-28.
141 engagement2 sourcesredditWed, Aug 26, 2026, 01:04 AM
Developments·Paper
Researchers introduce EvoUndo to verify recoverability of LLM agent self-modifications
An arXiv paper (2026-08-28) introduces EvoUndo, a framework for representing, synthesizing, diagnosing, and independently verifying whether model-generated self-modifications to LLM agent harnesses can be safely reversed. It targets the risk that capability-improving mutations to prompts, tools, middleware, and execution harnesses leave persistent effects that cannot be undone in different states. Across 600 unseen one-shot self-evolution tasks, researchers identified 197 capability-improving mutations that failed recoverability verification.
An arXiv paper (2026-08-28) introduces EvoUndo, a framework for representing, synthesizing, diagnosing, and independently verifying whether model-generated self-modifications to LLM agent harnesses can be safely reversed. It targets the risk that capability-improving mutations to prompts, tools, middleware, and execution harnesses leave persistent effects that cannot be undone in different states. Across 600 unseen one-shot self-evolution tasks, researchers identified 197 capability-improving mutations that failed recoverability verification.
1114 engagement3 sourcesarxiv, redditFri, Aug 28, 2026, 02:15 PM