arXiv:2510.26418v4 Announce Type: replace Abstract: Large Reasoning Models (LRMs) improve task performance through extended inference-time reasoning. Although previous studies suggest that longer reasoning should lead to more robust safety behavior, we find evidence to the contrary: over-extended reasoning can instead be exploited to systematically weaken refusal behavior. We propose Chain-of-Thought Hijacking, a simple yet effective black-box jailbreak attack that induces LRMs to engage in prolonged benign puzzle-solving reasoning, often lasting more than five minutes, before eliciting harmful compliance. Across HarmBench, CoT Hijacking achieves attack success rates of 99%, 94%, 100%, and 94% on Gemini 2.5 Pro, ChatGPT o4 Mini, Grok 3 Mini, and Claude 4 Sonnet, respectively. To understand why this attack succeeds, we conduct activation probing, attention-pattern analysis, and causal interventions on open-source reasoning models. Our results indicate that refusal behavior depends on a low-dimensional safety signal whose expression weakens as reasoning traces grow longer. In particular, extended benign reasoning shifts attention away from harmful intentions and attenuates refusal-related activations, producing what we call refusal dilution. These findings demonstrate that excessively prolonged reasoning can introduce a systematic jailbreak attack surface. We release our evaluation materials to support reproducibility and further research.

arXiv:2603.06603v1 Announce Type: cross Abstract: Data duplication during pretraining can degrade generalization and lead to memorization, motivating aggressive deduplication pipelines. However, at web scale, it is unclear what constitutes a ``duplicate'': beyond surface-form matches, semantically equivalent documents (e.g. translations) may induce redundant training signals once models become sufficiently capable. Practically, this means that semantic duplicates operate increasingly like exact duplicates during training. We present evidence that duplication is scale-dependent in two ways. First, as model capability increases, cross-entropy loss gradients for semantically equivalent documents become more aligned. Smaller models, by contrast, produce gradients that reflect surface similarity (e.g., shared tokens) rather than semantic similarity. Second, we embedded all 192 million FineWeb-Edu-Dedup documents using EmbeddingGemma-300m. For moderate corpus sizes, the cosine similarity between nearest-neighbors follows an isotropic power law baseline. However, as corpus size grows to hundreds of billions of tokens, the nearest-neighbor similarities deviate sharply, indicating accelerated semantic collisions. Finally, controlled pretraining on data sampled with replacement from pools of finite unique documents shows that limited uniqueness yields mild degradation for small models, but rapidly increasing loss penalties for larger models, breaking naive scaling extrapolation. We derive explicit scaling laws that allow practitioners to estimate deviation from expected scaling due to limited semantic uniqueness of the pretraining corpus. Our results identify and resolve an unstudied source of scale-dependence, allowing for more accurate prediction at scale.

arXiv:2603.06612v1 Announce Type: cross Abstract: Pass@k and other methods of scaling inference compute can improve language model performance in domains with external verifiers, including mathematics and code, where incorrect candidates can be filtered reliably. This raises a natural question: can we similarly scale compute to elicit gains in truthfulness for domains without convenient verification? We show that across five benchmarks and models, surprisingly, it cannot. Even at 25x the inference cost of naive sampling, polling-style aggregation yields no consistent accuracy gains over single-sample baselines and often amplifies shared misconceptions. We find that under uncertainty, models are better at predicting what other models will say within model ensembles than at identifying what is true, revealing a separation between social prediction and truth verification. Across models and benchmarks, aggregation fails to provide a robust truth signal because language model errors are strongly correlated. The source of correlation goes beyond any individual benchmark: we show that even when conditioned on out of distribution random strings and asked to produce pseudo-random outputs, different models produce correlated outputs. Confidence-based weighting provides no benefit because self-reported confidence fails to reliably distinguish correct from incorrect answers. These results delineate a boundary for inference-time scaling: in verified domains, additional samples provide more candidates for a verifier to filter; in unverified domains, additional samples merely reinforce shared misconceptions.