The paper investigates why vision‑language models that tokenize images with vector‑quantized (VQ) codebooks frequently hallucinate objects on grounded yes/no tasks. By applying activation patching across 25 models from eight large‑language‑model families, the authors uncover an early‑layer attention routing circuit shared by VQ‑tokenized VLMs. They develop a three‑gate diagnostic that isolates ten models carrying this circuit, show that swapping a single architectural component (VQ+Linear) introduces the circuit, and demonstrate that ablating the early‑layer ($L_0$) component reduces hallucinations in open‑ended generation by 31 % while other decoding‑time fixes do not.
By Shamanthak Hegde, Xiangrui Liu, Maitreya Patel, Yezhou Yang
arXiv:2606. 02628v1 Announce Type: new Abstract: We investigate whether open-source LLMs encode a linearly separable truthfulness signal in their hidden states, and at which network depth this signal is strongest.
By Aizierjiang Aiersilan
arXiv:2606. 00926v1 Announce Type: new Abstract: Mechanistic studies of sequence models often treat layerwise state encodings as architectural traits: recurrent models concentrate readable state, attention-based models distribute it.
By Yuhang Jiang
arXiv:2605. 07284v2 Announce Type: replace Abstract: A late-layer change learned during post-training may work on the base model's earlier state, or it may depend on earlier computation learned with it.
By Yifan Zhou
The paper introduces DARTS, a method for tuning decoder representations during model merging. It addresses representation bias in autoregressive decoders by using an entropy‑weighted L1 loss and a per‑position additive bias to correct errors that accumulate across token positions. Experiments on code generation, mathematical reasoning, and instruction following with Llama‑2‑7B show that DARTS improves performance over standard surgery while adding only 0.1% extra parameters.
By Aaryan Ajay Sharma, Sai Nishanth Padala, Seganrasan Subramanian
Optimal transport (OT) has been shown to detect hallucinations in neural machine translation (NMT) by measuring the geometric distance between cross-attention distributions and a reference distribution, without any supervision. We extend this analysis to all six decoder layers of the Fairseq DE-EN model ($N=3{,}414$), showing that Wass-to-Unif and Wass-to-Data are complementary detectors specialised across hallucination types, that detection is concentrated in layers L1--L4 with L5 anti-predictive for subtler types, and that hallucinated translations lack the exploratory attention phase present in correct translations from the first decoding step.