Post-trained LLM 如何認出自己在說話？
IAS 與 Anthropic 研究者的論文發現，Post-trained 模型生成自身回答時的熵值比閱讀外部文本低 3-4 倍。本文探討「輸入驚喜度」如何驅動這種自我認知機制，並分析其對 AI Agent 穩定性的影響。
Anthropic
How Post-trained LLMs Recognize Their Own Voice
A paper by researchers at IAS and Anthropic finds that post-trained models’ output-distribution entropy is 3–4 times lower when generating their own responses than when reading external text. This article explores how Input Surprise drives this self-recognition mechanism and analyzes its implications for AI agent stability.
當模型自信地輸出錯誤：價值軸線如何編碼生成軌跡
Anthropic Fellows 研究在 Qwen3-8B 中發現一條線性方向，其投影與信心、回溯及程式碼正確性相關。理解這條軸線與 DPO 訓練的交互作用，有助於評估 Eval Awareness 風險與模型行為邊界。
When Models Confidently Produce Errors: How the Value Axis Encodes Generation Trajectories
Anthropic Fellows research found a linear direction in Qwen3-8B whose projection is associated with confidence, backtracking, and code correctness. Understanding how this axis interacts with DPO training helps assess Eval Awareness risks and the boundaries of model behavior.
SAE 特徵提取：從殘差流看見模型內部表示
Anthropic 的研究團隊在 Claude 3 Sonnet 中訓練了高達 3400 萬個特徵的稀疏編碼器（SAE），提取出可解釋的特徵表示。研究顯示能對模型輸出行為產生因果影響，為 AI 安全治理提供了可研究的內部訊號與潛在工具，但忠實度仍需驗證。
SAE Feature Extraction: Seeing Model Representations Through the Residual Stream
Anthropic’s research team trained sparse autoencoders (SAEs) with up to 34 million features in Claude 3 Sonnet, extracting interpretable feature representations. The research shows causal effects on model output behavior, providing internal signals and potential tools that can be studied for AI safety governance, though faithfulness still requires validation.
Constitutional AI: 如何透過 AI Feedback 實現自動化安全治理
標註員 A 給 5 分，B 給 1 分。同一個回答。傳統 RLHF 的瓶頸不在成本，而在訊號雜訊比。本文解析 Constitutional AI 如何透過 AI Feedback 建立自動化安全邊界，以及落地時需要評估的成本結構。
Constitutional AI: How AI Feedback Enables Automated Safety Governance
Annotator A gives it a 5, annotator B gives it a 1. The same answer. The bottleneck in traditional RLHF is not cost, but the signal-to-noise ratio. This article breaks down how Constitutional AI uses AI Feedback to build automated safety boundaries, and the cost structure teams need to evaluate when putting it into practice.
Toy Models of Superposition：神經網路的特徵壓縮術
打開 IDE。你觀察隱藏層的單一神經元。它同時響應了 ‘Python’ 和 ‘錯誤處理’。這不是 bug，這是特徵壓縮。Anthropic 的 Toy Models 研究揭示特徵稀疏性如何驅動「疊加（Superposition）」現象，解釋模型如何在容量限制下儲存更多概念。
Toy Models of Superposition: Feature Compression in Neural Networks
Open your IDE. You observe a single hidden-layer neuron. It responds to both “Python” and “error handling.” This is not a bug; it is feature-level compression. Anthropic’s Toy Models research reveals how feature sparsity drives “superposition,” explaining how models store more concepts under capacity constraints.