…Error analysis shows that moderation relations and concept hierarchies remain the most challenging cases, highlighting the difficulty of extracting higher-order empirical claims and implicit abstraction structure from scientific abstracts. View arXiv…
…it's refreshing to see upstream recognition and autograding evaluated together, and the claim that fewer than 5% of cases need human regrading to get near-expert performance feels realistically deployable. i…
…into a single model, identifying capability loss in different ways: mixed RLVR suffers from inter-capability divergence cost, while the pipeline of first training experts and then performing OPD , though avoiding divergence…
…Travis Lelle Abstract LoRA adapters can be backdoored through training data poisoning while maintaining performance, with the backdoor activating at token feature level and being detectable through behavioral and weight-level statistics…
…Xin Li , , , , Abstract On-policy distillation with reward extrapolation exhibits a safety threshold beyond which structured output tasks lose format preservation, with empirical validation showing performance parity at reduced parameter count when…
…We do not propose a new rewriter or mitigation; we release the intervention runner and the sentinel panel so that other rewriter-gain claims can be tested against the same standard. View…
…the model must decide how to search while also remembering what it has seen, which evidence is useful, which constraints remain open, and which claims have actually been checked. We argue that…
…performance of Qwen2.5-Math-Instruct and Qwen2.5-7B-Instruct to quantify the improvement using reasoning trace SFT distillation. How exactly is the Qwen/Qwen2.5-Math-RM-72B model used…
…models with an “agentic framework” that can autonomously browse the web, gather info, and synthesize structured answers. I appreciate how you walk through the rationale behind agent frameworks, discuss the performance metrics…
…Second, an augmentation baseline equipped with test-time orbit averaging matches the equivariant model exactly -- bit-identical per-epoch validation curves across matched cells -- so the architecture-vs-augmentation gap is conditional…
To show you the most relevant results, we’ve omitted some entries very similar to those already shown. Repeat the search with the omitted results included.