Research map / Large language models / Reasoning and chain of thought

Beyond External Monitors: Enhancing Transparency of Large Language Models for Easier Monitoring

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TELLME, which fine-tunes LLMs with LoRA using a contrastive (InfoNCE) loss over hidden representations to cluster same-behavior and separate different-behavior activations, making models easier to monitor and improving their safety without external modules.

Problem

LLM internal thinking is opaque and CoT does not faithfully reflect it; existing representation-based monitoring relies on external modules (SAEs, transcoders) applied to unaltered representations, making monitoring passive and constrained by entangled features.

Approach

Directly modify the LLM's own weights with a contrastive disentanglement loss on intermediate representations (positive pairs from same behavior, negatives from different behaviors) plus a retain loss, so the model's representations are inherently more separable for monitors and detoxification improves without preference data.

Key results

• On Llama-3.1-8B, TELLME raises Self-Sim monitoring accuracy on Multi-Risk from 59.2% to 73.0% and Safe-or-Harmful from 73.4% to 82.5%.
• Average safety on BeaverTails and SaladBench improves by 8.1 points over the original model, while average capability stays near unchanged (e.g., Llama-3.1-8B 67.1 to 65.8).
• TELLME with 41M LoRA parameters outperforms an SAE with ~1.09B parameters on Multi-risk and binary monitoring accuracy (83.5% vs 76.6% on binary classification).

Surprising findings

• Applying TELLME to Qwen2.5-7B improves safety while standard SFT on refusal data drops its safety from 92.1 to 58.4 on BeaverTails.
• The TELLME improvement on Qwen3-VL models grows with model scale (235B gains +5.8 multimodal safety).

Limitations

• Monitoring gains are mixed for the judge-based monitor, e.g., Mistral-7B Judge on Multi-Risk drops relative to origin in some variants.
• Comparison with SAE baselines is acknowledged as not strictly fair due to differing training paradigms and scales.
• Safety judgments rely on an LLM judge (Qwen3-Guard-Gen-8B), whose errors can affect absolute scores.

Experimental setup

Datasets: BeaverTails UltraChat SPA-VL RLHF-V MATH MMLU GLUE CoLA SiQA

Benchmarks: BeaverTails SaladBench XSTest GSM8K MMLU AGIEval MM-Safety VLSBench MSSBench MMMU MMLU-Pro CoLA

Metrics: monitoring accuracy safety rate refusal rate average capability score coding rate eRank l2 distance angle Hausdorff distance

Baselines: SAE SAE-KL SAE-Multi-TopK Transcoder Skip Transcoder SFT Self-Sim Linear Probe

Builds on: InfoNCE NT-Xent Sparse Autoencoders Transcoders LoRA

Closest papers by meaning

• Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.88)
• CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.87)
• Tuning-Free Accountable Intervention for LLM Deployment – a Metacognitive Approach (AAAI 2025 · similarity 0.87)
• Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.87)
• Can large language models explore in-context? (NeurIPS 2024 · similarity 0.87)
• LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · similarity 0.87)