The paper introduces Feedback Distillation, a novel training method that uses a language model's privileged feedback to provide token-level supervision, significantly improving complex reasoning tasks…

The paper proposes Skill-Conditioned Gated Self-Distillation (SGSD), a novel framework that uses retrieved, potentially noisy skills to guide LLM reasoning, achieving state-of-the-art performance on m…

DenseSteer is a training-free inference-time framework that improves the math reasoning capabilities of small language models by steering their internal representations toward a 'Dense Reasoning' patt…

The OISD framework improves language model reasoning by distilling on-policy predictive signals from the final output layer to intermediate representations, leading to substantial improvements on math…

The paper introduces Contrastive Reflection (CORE), a novel non-parametric method that rapidly improves language model reasoning by distilling contrasts between successful and unsuccessful problem att…

ThinkSwitch introduces a low-compute co-training procedure that distills the reasoning benefit of large language models into weights, significantly improving performance on specific reasoning tasks.

The paper introduces Trajectory-aware OPD (TOPD), a method that uses near-future trajectory information to improve On-Policy Distillation by accurately identifying and guiding true reasoning divergenc…

The paper analyzes the failure modes of aggressive 2-bit quantization in large reasoning models, proposing lightweight controls like FP16 planning and loop rescue to restore accuracy and achieve pract…

The paper argues that using confidence-based decoding, which is optimized via training mask alignment, fundamentally misaligns Masked Diffusion Models (MDMs) from the logical flow needed for complex r…

The paper introduces Canonical-Context On-Policy Distillation (CCOPD) to improve multi-turn language model performance by mitigating 'self-anchored drift,' ensuring consistent answers regardless of wh…

The paper introduces Lookahead Group Reward (&) to combat Supervision Fidelity Decay (SFD) in on-policy distillation, significantly improving student model performance on long reasoning tasks.

The paper introduces COLAGUARD, a novel guardrail model that efficiently transfers multi-step safety reasoning into a continuous latent space, achieving state-of-the-art safety performance with massiv…

The paper introduces COLAGUARD, a novel guardrail model that efficiently transfers multi-step safety reasoning into a continuous latent space, achieving high safety performance with massive improvemen…

Weak self-training on synthetic data can amplify a language model's existing capabilities, but this effect is strictly dependent on the compatibility between the source and student models, not on the…

The paper introduces an automatic numeric-remapping attack to test the robustness of LLMs on arithmetic word problems, finding that LLMs remain sensitive to small numeric changes in datasets like GSM8…

The paper introduces Chunk-Level Guided Generation, a training-free method that uses an off-the-shelf large language model (LLM) as a process scorer to guide small model generation, achieving performa…

The paper introduces Lossless Anti-Distillation Sampling (LADS), a novel sampling scheme that makes harvested data correlated for malicious distillers while ensuring benign users receive statistically…

The paper introduces TraceGuard, a detectability-aware antidistillation method that identifies and poisons 'thought anchors'—sparsely critical sentences—to degrade student model learning without makin…

The paper proposes EKSFT, a selective fine-tuning method that masks high-entropy or high-KL divergence tokens during Supervised Fine-Tuning (SFT) to prevent distribution shift and improve subsequent R…