Multimodal models have often scored below text-only baselines in the BabyLM Challenge (Klerings et al., 2024; Takmaz et al., 2025). We test whether additional recurrent computation helps close this gap. Recurrence increases computation by reusing layers without adding parameters or training data. Using a single implementation, we compare standard and looped Transformers under a 100M-word budget, with three seeds per main condition. Each image is represented by one feature vector from a frozen DINOv2 encoder, used as a prefix token. Caption words count toward the budget, so multimodal training also changes the language data. In a pre-registered text-only comparison, a 12-layer stack repeated four times exceeds a standard 48-layer Transformer by 5.4 BLiMP percentage points at matched inference compute. However, multimodal training lowers BLiMP accuracy by about 3.8 points in both architectures, with no clear mitigation from recurrence. An exploratory loop-side no-image control, trained with the multimodal language data and batching schedule, scores 5.7 points below the text-only baseline. Adding the visual prefix raises accuracy by 1.85 points on average, but this contrast is uncertain with three seeds. These results suggest that changes in language data and its presentation are a major source of the drop. Diagnostic tests show that the models use visual information; its contribution to grammatical generalization remains uncertain under this single-vector interface.
Looped Transformers extend inference-time computation through repeated latent-state updates, but final-depth training does not directly encourage useful intermediate alternatives. We introduce Best-of-Depth (BoD), which trains a recurrent trajectory as a candidate set by minimizing prediction loss across depths. Different valid targets can be matched at different depths without architectural changes or depth-specific labels. An exact-match analysis connects this objective to solution coverage. On N-Queens completion and graph coloring, BoD improves coverage over final-depth and intermediate-supervision objectives. N-Queens coverage reaches 91.03%, versus 45.75% for final-depth training. Further experiments show discovery beyond the training depth and recovery of solutions withheld from teacher pools on familiar inputs. Fine-tuning a pretrained looped language model with BoD yields higher pooled code-generation success than final-depth training when programs are generated separately at evaluated fixed depths. These results support training recurrent computation for solution discovery along its trajectory.
Clinical decisions are often required under incomplete information. Clinical experts must identify whether available information is sufficient for judgment, as both premature conclusions and unnecessary abstention can compromise patient safety. To evaluate this capability of large language models (LLMs), we developed ClinDet-Bench, a benchmark based on clinical scoring systems that decomposes incomplete-information scenarios into determinable and undeterminable conditions. Identifying determinability requires considering all hypotheses about missing information, including unlikely ones, and verifying whether the conclusion holds across them. We find that recent LLMs fail to identify determinability under incomplete information, producing both premature conclusions and excessive abstention, despite correctly explaining the underlying scoring knowledge and performing well under complete information. These findings suggest that existing benchmarks are insufficient to evaluate the safety of LLMs in clinical settings. ClinDet-Bench provides a framework for evaluating determinability recognition, leading to appropriate abstention, with potential applicability to medicine and other high-stakes domains, and is publicly available.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Gouki Minegishi, Hiroki Furuta Google DeepMind, Takeshi Kojima The University of Tokyo, Yusuke Iwasawa The University of Tokyo, The University of Tokyo, Yutaka Matsuo The University of Tokyo
Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a mechanistic account based on the geometry of feature superposition. Because features are encoded in overlapping, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this mechanism and empirically test it across multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, gpt-oss 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach—filtering training samples nearest to toxic features—reduces misalignment by 34.5%, substantially outperforming random removal and achieving stronger mitigation than LLM-as-a-judge–based filtering. Our study explains emergent misalignment through feature superposition, providing a basis for understanding and mitigating this phenomenon.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表