Tag
A year-long adversarial verification layer over an agent stack found that roughly 3 in 10 confident outputs failed review, arguing that model agreement is not verification and that failure-cost auditing matters more than consensus.
论文提出 MOTIVE 框架，通过多视角自验证学习可靠性分数，在多模态推理中做出接受或重新思考的决策，从而提升视觉语言模型答案的可靠性并减少不必要的推理轮次。
Salesforce AI Research 提出 CLIFT，一种基于一致性自验证的网页智能体训练与测试时扩展方法，使 31B 开源 Gemma-4 智能体在 WebArena Infinity 上达到 74.6%，超越使用浏览器的 Gemini 3 Flash，并且在部署时无需调用前沿模型评判器。
CLIFT 利用共形自验证(conformal self-verification)将稀疏的judge反馈转化为可复用的训练奖励与无需外部judge的测试时轨迹选择，在WebArena Infinity、VisualWebArena和Online Mind2Web上取得了开源web agent的最优表现。
MOTIVE提出一个让视觉-语言模型自主判断何时信任答案、何时重新思考的框架，通过五个互补的验证视角和学习到的可靠性分数，在五个基准上将Qwen2.5-VL模型平均准确率提升约8个百分点。
This paper introduces a diagnostic framework for terminal agents' self-verification on TerminalBench2.1, finding that agents rarely detect or repair errors (only 61.43% of incorrect candidates detected, 49.36% of those repaired), and proposes Student-Conditioned Verification Distillation (SCVD), which improves Qwen3.5 Pass@1 by 9.74–16.85 points while avoiding out-of-distribution degradation on SWE-bench Verified.
The article discusses the transformation in AI applications towards long-range closed-loop self-verification, highlighting agent workflows from Space Bunny and Google Research that enable autonomous testing and error correction, reshaping development roles.
LLM-as-a-verifier is a framework providing fine-grained feedback for AI agents, achieving state-of-the-art performance on benchmarks like Terminal-Bench 2.1 with DeepSeek V4 Flash, outperforming Claude Fable 5 at lower cost.
A developer describes letting their AI agent autonomously test its own model upgrade by running controlled probes and measuring performance, revealing issues that throttled itself.
An AI agent built a knowledge graph from a research paper it had never seen before, using self-verification techniques to reduce hallucinations.
This paper presents the first comprehensive overview of metacognition in LLMs, arguing that behaviors like confidence calibration and self-verification are facets of a unified metacognitive ability, and taxonomizes methods and benchmarks for evaluating and improving these abilities to enhance LLM reliability and transparency.
Introduces SVR-R1, a multi-turn reinforcement learning framework that uses the model's own verification as a learning signal for multi-modal reasoning, achieving significant accuracy improvements over standard GRPO baselines on vision-language reasoning benchmarks.
Apodex-1.0-H is a new deep research model that introduces a multi-agent architecture where the model decomposes tasks, spawns specialist sub-agents, and uses self-verification and iterative improvement to produce answers. Open-weight variants are available on HuggingFace.
A 4 billion parameter open model from the Apodex family outperforms 30 billion parameter models on web research benchmarks, attributed to careful training data and self-verification techniques rather than raw scale, suggesting a more democratic trajectory for AI capability.
Discussion on the importance of self-verification loops in AI models like Claude to improve reliability and reduce the need for manual oversight.
FineVerify is a self-verification framework for agentic search that decomposes questions into sub-questions, verifies sampled candidates, and selects the best one, achieving substantial accuracy improvements over baselines on multiple benchmarks, including enabling GPT-5-mini to surpass GPT-5 on BrowseComp-Plus.
Proposes Self-Verified Distillation, a method where LLMs generate and self-verify candidate solutions from unlabeled seed questions using prompt-based verification, then train on the filtered dataset, achieving significant gains on math, science, and coding benchmarks across Qwen3 models.
This paper investigates how large reasoning models can detect and correct their own errors internally, identifying a highly interpretable critique vector that enhances error detection without additional training, improving test-time scaling performance.
RLanceMartin highlights new self-verification (Outcomes) and self-learning (Dreaming) features for Claude discussed at the Code With Claude event.