Study tests Jacobian Lens entropy as error predictor on Qwen3-4B across 7 datasets
A researcher evaluated whether entropy in a language model's internal 'workspace' (inspired by Anthropic's Jacobian Lens) can predict confidently incorrect answers. Testing Qwen3-4B on ~11,400 examples from seven datasets, they found workspace entropy can complement output confidence for error routing on factual retrieval tasks like PopQA.
Entities
Related
Community stress-tests Anthropic's J-Space hallucination signal on Qwen3-4B across 7 datasets
A developer mapped Anthropic's J-Space hallucination detection method onto Qwen3-4B, testing it across ~11,400 examples from 7 datasets. The work builds on Anthropic's paper and an open-source implementation by solarkyle, aiming to assess whether internal workspace entropy is a deployable hallucination detector.
↑ Updated Sun, Jul 12, 2026, 05:06 AM — Stress-test results across 7 datasets posted on Reddit.
Researchers identify asymmetric generalization problem in LLM unlearning benchmarks
A new arXiv paper argues that existing machine unlearning benchmarks for LLMs suffer from under-forgetting and over-forgetting due to an asymmetric generalization problem. The authors propose that evaluation must cover diverse query formulations of target facts to reliably measure knowledge removal while preserving unrelated capabilities.
Community applies Anthropic's J-space lens to open Qwen3-8B for agent guardrails
A Reddit user replicated Anthropic's J-space research on the open-source Qwen3-8B model, using a Jacobian lens to detect silent internal reasoning before tool calls. They then wired this into agent guards to intercept prose drift (e.g., leaning toward natural language instead of JSON), enabling stop/cancel/keep decisions.
Anthropic reveals J-space inside Claude Opus 4.6, offering clearest view yet of LLM internal reasoning
Anthropic developed the Jacobian lens (J-lens) to uncover a hidden area called J-space inside Claude Opus 4.6, revealing individual words related to the model's output. Separately, a researcher released the BABEL codec for GPT-2 small, achieving 94.7% behavior reconstruction and enabling bidirectional reading/writing of the model's internal state. These advances give practitioners unprecedented insight into how language models think.
VLMs encode correct object counts internally but output wrong answers, study finds
A new arXiv paper investigates why vision-language models (VLMs) fail at basic object counting. By training probes on internal activations across four VLMs and five counting datasets, researchers found that nonlinear probes can reliably detect counting errors, indicating that models often encode the correct count even when they output the wrong answer. SVCCA analysis further confirms a misalignment between internal representations and verbalized outputs.