arXiv AI

The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining

arXiv Machine Learning
Sep 17

Attention Dispersion as a Diagnostic Signal for Hallucination in Large Language Models

The paper proposes using the temporal volatility of internal attention mechanisms—measured by an unsupervised attention dispersion metric—as a diagnostic signal for hallucinations in large language models. It demonstrates that spikes in attention entropy within intermediate layers correlate with reasoning breakdowns, and shows statistically significant AUC improvements of up to +0.076 over output-based baselines on GSM8K and MATH-500 benchmarks using the Qwen2.5 model family.

By Shardul P. More, Tanuja S. Pawar
arXiv Computation and Language
Aug 28

Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update

The paper investigates how large language models exhibit sycophancy—changing answers to align with user feedback—and distinguishes two types of answer flips: Unsupported‑Yielding (merely satisfying the user) and Rational‑Updating (truly incorporating useful evidence). Using a two‑turn evaluation framework, the authors show that anti‑sycophancy methods often trade off between reducing Unsupported‑Yielding and preserving Rational‑Updating, even when both objectives are jointly optimized. Mechanistic analysis reveals overlapping neural substrates for the two behaviors, suggesting that effective interventions should focus on selective suppression rather than blanket suppression.

By Huanhuan Ma, Henry Peng Zou, Chengze Li, Enze Ma, Yunyue Su, Philip S. Yu
arXiv AI
Sep 3

Language Models Can Control Their Own Attention

The paper introduces Declarative Attention (DA), a protocol that lets language models explicitly declare which parts of their context to focus on during generation. By partitioning decoding into full-context, region-specific, and recent-output-only modes, the inference engine can skip large portions of the KV cache, dramatically reducing attended tokens. Experiments on 15 long-context tasks with off-the-shelf models show significant savings (52.0% and 31.1% reductions) with only modest accuracy drops that diminish as model size increases.

By Namgyu Ho, Huzama Ahmad, Woosung Koh, Se-Young Yun, Tal Schuster, Cicero Nogueira dos Santos
arXiv AI
6d ago

From Architecture to Output: Testing the Structural Origins of Hallucination and the Amplifying Role of Data

The paper investigates whether hallucinations in language models stem from their architecture or from the data they are trained on. By swapping attention mechanisms for state‑space mixers in Pythia/Mamba models, the authors show that attention models are more prone to repeating false statements, while state‑space models better handle planted facts. A second study demonstrates that correcting a single erroneous fact in instruction‑tuned models significantly reduces downstream errors, indicating that the structure of the model influences how errors propagate.

By Md. Rejaul Korim Sadi, Toufiqur Rahman Tasin, Golam Mostofa Naeem
arXiv Computation and Language
Sep 4

How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits

The study investigates the consistency and specificity of language model circuits across six tasks and five models, focusing on component-level (attention heads and MLP blocks) and neuron-level circuits. Component-level circuits are highly consistent and causally important but lack task specificity, as ablating a circuit for one task similarly harms performance on other tasks. Neuron-level circuits show higher task specificity but lower consistency, with overlap mainly between closely related tasks. The analysis of Llama‑3.2‑3B reveals that shared components are predominantly MLP blocks, while attention heads act as generic attention‑sink heads.

By Michael Li, Nishant Subramani
arXiv AI
Oct 12

Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs

The paper investigates how safety‑aligned large language models (LLMs) can still be tricked by jailbreaks or over‑refuse benign requests. By analyzing residual‑stream updates in decoder‑only Transformers, the authors find that harmfulness is more readily decoded from attention updates—especially in early layers—while the MLP component is primarily responsible for writing refusal responses. Post‑training enhances MLP‑driven refusal writing without improving harmfulness decodability, and causal interventions confirm that downstream MLP updates are crucial for the refusal shift caused by upstream attention changes. Building on these insights, the authors propose Sentinel, a system that combines an attention‑derived perception score with an MLP‑derived decision score to better detect jailbreaks and over‑refusals across multiple models.

By Zixuan Weng, Haoqi He, Weifei Jin, Yue Huang, Yang Bai