ALEE generates AMR-based English minimal pairs with fine-grained semantic shifts, translates them, and evaluates embedding models on 275+ languages to expose cross-lingual gaps linked to training data and tokenization.
Transactions of the Association for Computational Linguistics , volume=
7 Pith papers cite this work. Polarity classification is still indexing.
7
Pith papers citing it
years
2026 7verdicts
UNVERDICTED 7representative citing papers
LLM chain-of-thought crosses a commitment boundary early; subsequent steps are epiphenomenal, enabling early-exit that shortens traces 55% with negligible performance change.
ACTS models LLM reasoning control as an MDP solved by a controller agent initialized on synthetic multi-budget trajectories and refined with budget-conditioned RL, achieving token savings while matching full-reasoning accuracy.
SEMJ is a self-evolving multilingual LLM judge that turns cross-lingual inconsistency into iterative self-reflection, outperforming voting and reflection baselines on accuracy and consistency.
The paper introduces a three-source decomposition showing that answer flips in multi-agent LLM debate include 37% spontaneous instability and 29% harmful conformity, with even vacuous reasoning persuading 20-39% of resistant agents and interventions reducing harmful conformity by 13.6 points.
Dutch LLMs display coherence illusions tracked by surprisal, with attention entropy identifying affected heads and a new energy metric quantifying discourse coherence.
Merging any combination of monolingual pre-trained models leads to performance collapse due to interference, indicating that merging flexibility from fine-tuning does not extend to pre-training.
citing papers explorer
-
ALEE: Any-Language Evaluation of Embeddings via English-Centric Minimal Pairs
ALEE generates AMR-based English minimal pairs with fine-grained semantic shifts, translates them, and evaluates embedding models on 275+ languages to expose cross-lingual gaps linked to training data and tokenization.
-
Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning
ACTS models LLM reasoning control as an MDP solved by a controller agent initialized on synthetic multi-budget trajectories and refined with budget-conditioned RL, achieving token savings while matching full-reasoning accuracy.
-
When Languages Disagree: Self-Evolving Multilingual LLM Judges
SEMJ is a self-evolving multilingual LLM judge that turns cross-lingual inconsistency into iterative self-reflection, outperforming voting and reflection baselines on accuracy and consistency.
-
Not All Flips Are Conformity: Decomposing Stance Convergence in Multi-Agent LLM Debate
The paper introduces a three-source decomposition showing that answer flips in multi-agent LLM debate include 37% spontaneous instability and 29% harmful conformity, with even vacuous reasoning persuading 20-39% of resistant agents and interventions reducing harmful conformity by 13.6 points.
-
When Context Misleads: Surprisal, Energy and Attention Entropy as Metrics of Coherence Illusions in LLMs
Dutch LLMs display coherence illusions tracked by surprisal, with attention entropy identifying affected heads and a new energy metric quantifying discourse coherence.
-
On the Limits of Model Merging for Multilinguality in Pre-Training
Merging any combination of monolingual pre-trained models leads to performance collapse due to interference, indicating that merging flexibility from fine-tuning does not extend to pre-training.