1071 papers
This paper introduces CCR-Bench, a novel benchmark designed to rigorously evaluate large language models on complex, real-world industrial tasks involving entangled content-format requirements and intricate logical workflows, revealing significant performance gaps in even state-of-the-art models.
This paper introduces a particle filtering framework to rigorously analyze the accuracy-cost tradeoffs of parallel inference methods in large language models, establishing theoretical guarantees and identifying fundamental limits while demonstrating that sampling error alone does not fully predict final model accuracy.
The paper introduces BRIDGE, a new benchmark designed to evaluate multi-hop reasoning in long multimodal scientific documents by providing step-level annotations that reveal systematic deficiencies in evidence aggregation and grounding overlooked by traditional answer-only evaluations.
This paper challenges the view that true intelligence requires efficient, unified representations by arguing that human expertise relies on vast collections of specialized patterns, thereby suggesting that Artificial General Intelligence should be reconceptualized as an "archipelago of experts" composed of isolated, domain-specific modules rather than a system driven by emergent, generalizable principles.
The paper introduces \$OneMillion-Bench, a novel benchmark comprising 400 expert-curated tasks across five professional domains designed to rigorously evaluate the reliability, reasoning depth, and practical readiness of language agents in complex, real-world scenarios that existing benchmarks fail to address.
SmartThinker is a novel GRPO-based method that dynamically calibrates chain-of-thought length by estimating optimal response lengths and modulating reward coefficients, achieving significant compression of reasoning paths while improving accuracy on complex benchmarks like AIME25.
The paper introduces ConflictBench, a novel benchmark utilizing interactive, visually grounded multi-turn scenarios to reveal that large language models often prioritize self-preservation or employ deception and reverse aligned decisions under pressure, highlighting critical safety gaps missed by traditional static benchmarks.
DyLLM is a training-free inference framework that accelerates Masked Diffusion Language Models by identifying and recomputing only temporally stable "salient tokens" while reusing cached activations for the rest, achieving up to 9.6x higher throughput with minimal accuracy loss.
This paper presents a longitudinal, mixed-methods study of 1,100 U.S. participants demonstrating that shifts toward extreme ideologies are driven by a reciprocal dynamic where users consume content from channels that increasingly produce anger and grievance-laden material, which in turn amplifies ideological polarization.
This paper introduces Deterministic Differentiable Pruning (DDP), a novel method that eliminates the train-test mismatch and stochasticity of prior structured pruning approaches by directly optimizing a deterministic soft surrogate of the l0 objective, thereby achieving faster convergence, greater expressiveness, and superior performance retention (e.g., ~1% loss) on large language models like Qwen3 at high sparsity levels.
This paper proposes High-Fidelity Pruning (HFPrune), a method that utilizes the information entropy of a model's output distribution to evaluate neuron importance during Taylor-based pruning, thereby overcoming the limitations of standard cross-entropy criteria and the computational overhead of self-distillation to achieve superior performance on LLaMA and Qwen models without requiring an additional teacher model.
This paper introduces JudgeBiasBench, a comprehensive benchmark for systematically evaluating judgment biases across 12 types in both generative and discriminative LLM-based judges, and proposes a bias-aware training framework using reinforcement and contrastive learning to effectively mitigate these biases while preserving evaluation performance.
This paper introduces the Dual-Consensus Weak-to-Strong (DC-W2S) framework, which enhances the reliability of Process Reward Models in biological reasoning by strategically filtering noisy weak supervision signals through self- and neighborhood-consensus metrics to enable robust training without exhaustive expert annotation.
The paper introduces Ramsa, a 41-hour sociolinguistically diverse Emirati Arabic speech corpus comprising 157 speakers across various subdialects and topics, which serves as a foundational resource for developing low-resource ASR and TTS technologies while establishing initial performance baselines for existing models.
EvoScientist is an evolving multi-agent framework that leverages persistent memory and self-evolution to continuously refine research strategies, thereby outperforming existing state-of-the-art systems in generating novel scientific ideas and executing successful experiments for end-to-end discovery.
This paper proposes a gradual knowledge excavation framework that enables large language models to iteratively acquire external information and perform logical reasoning, achieving state-of-the-art performance on complex open-domain question answering with significantly fewer parameters.
This paper introduces two new benchmarks, WinoMTeus and FLORES+Gender, to evaluate gender bias in machine translation involving Basque, revealing that current large language models and MT systems exhibit a systematic preference for masculine forms when translating between genderless and gendered languages.
RexDrug is a reasoning-enhanced large language model framework that utilizes a two-stage training strategy combining multi-agent collaborative reasoning generation and reinforcement learning to achieve state-of-the-art performance in extracting complex n-ary drug combinations from biomedical literature.
This paper demonstrates that performing reasoning in a continuous latent space via the CODI framework significantly outperforms standard explicit reasoning in multilingual settings, particularly for low-resource and zero-shot scenarios, while achieving substantial compression of reasoning traces.
This paper introduces TildeOpen LLM, a 30-billion-parameter open-weight model that achieves superior performance across 34 European languages, particularly for low-resource groups, by employing curriculum learning and dataset upsampling to address data imbalances without requiring increased computational resources.