Large reasoning models achieve strong problem-solving performance, but often generate excessively long chain-of-thought outputs, which substantially increase inference cost and limit practical deployment. Existing reasoning compression methods reduce length only partially and often suffer from unstable optimization or accuracy degradation. We trace this limitation to an underlying trajectory-level objective conflict: when the compression objective and the accuracy objective are optimized jointly, compression favors short-and-correct trajectories while accuracy optimization continues to preserve long-but-correct ones that remain useful on hard problems. This conflict creates opposing gradients on the same behaviors and can trap training in a local trade-off where further compression harms accuracy while accuracy recovery reintroduces verbose reasoning. At the model-decision level, we refer to the resulting failure mode as entropy conflict: the policy fails to fully collapse onto concise trajectories because accuracy-oriented updates keep restoring broader support over long-correct reasoning paths. To address this issue, we propose an entropy-guided training framework. Instead of forcing a mixed objective to follow an unstable path toward the final efficient-output set, we first contract the policy to a short-output set and then expand from this compact regime to recover accuracy and robustness. Experiments on multiple mathematical reasoning benchmarks show that our method achieves performance comparable to the original model using only about 25% of its reasoning tokens.

In this paper, we describe our submission for the NLI4CT 2024 shared task on robust Natural Language Inference over clinical trial reports. Our system is an ensemble of nine diverse models which we aggregate via majority voting. The models use a large spectrum of different approaches ranging from a straightforward Convolutional Neural Network over fine-tuned Large Language Models to few-shot-prompted language models using chain-of-thought reasoning.Surprisingly, we find that some individual ensemble members are not only more accurate than the final ensemble model but also more robust.

Recent multilingual pretrained language models (mPLMs) have been shown to encode strong language-specific signals, which are not explicitly provided during pretraining. It remains an open question whether it is feasible to employ mPLMs to measure language similarity, and subsequently use the similarity results to select source languages for boosting cross-lingual transfer. To investigate this, we propose mPLM-Sim, a language similarity measure that induces the similarities across languages from mPLMs using multi-parallel corpora. Our study shows that mPLM-Sim exhibits moderately high correlations with linguistic similarity measures, such as lexicostatistics, genealogical language family, and geographical sprachbund. We also conduct a case study on languages with low correlation and observe that mPLM-Sim yields more accurate similarity results. Additionally, we find that similarity results vary across different mPLMs and different layers within an mPLM. We further investigate whether mPLM-Sim is effective for zero-shot cross-lingual transfer by conducting experiments on both low-level syntactic tasks and high-level semantic tasks. The experimental results demonstrate that mPLM-Sim is capable of selecting better source languages than linguistic measures, resulting in a 1%-2% improvement in zero-shot cross-lingual transfer performance.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model’s diverse response styles such as starting with “Sure” or refusing to answer. Consequently, first-token evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.