6arXiv cs.LG (Machine Learning)·
Distill to Detect (D2D): surfacing stealth preferential biases in LLMs via cartridge distillation
Researchers introduce Distill to Detect (D2D), a method for exposing hidden preferential biases in language models by distilling the distributional shift between a suspected model and its base into a KV-cache prefix adapter (cartridge), amplifying the bias signal into generated text. The work addresses a fundamental asymmetry in bias detection: stealth biases introduced anywhere in a model's supply chain can remain invisible to text-based inspection, weight analysis, or representation probing. D2D is grounded in a theoretical framework based on Fisher-weighted projection of logit distribution shifts and is validated across multiple bias types. The technique offers a practical auditing tool for detecting hidden behaviors in deployed LLMs.
Related guides (2)
Related events (8)
DOPD: Advantage-aware dual on-policy distillation to address privilege illusion in LLM/VLM training
Researchers introduce DOPD (Dual On-policy Distillation), a knowledge distillation framework that dynamically routes token-level supervision between a privileged teacher and privileged student policy based on advantage gap and relative probabilities. The method addresses a failure mode called 'privilege illusion,' where information asymmetry between teacher and student is conflated with a transferable capability gap. Experiments on both LLM and VLM settings show DOPD outperforms vanilla on-policy distillation and related methods, with additional gains on stability, continual learning, and out-of-distribution tasks.
LLMSurgeon: Post-Hoc Auditing of LLM Pretraining Data Mixtures
LLMSurgeon formalizes Data Mixture Surgery (DMS), a framework for estimating the domain-level distribution of an LLM's pretraining corpus using only generated text from the target model. The method casts DMS as an inverse problem under the label-shift assumption, using a calibrated soft confusion matrix to correct domain confusion and recover the latent mixture prior. The authors also introduce LLMScan, a verifiable evaluation suite built from open-source LLMs with known pretraining mixtures, on which LLMSurgeon demonstrates high-fidelity recovery of domain compositions without access to training data.
Future Confidence Distillation enables pre-solution confidence estimation in LLMs
Researchers introduce 'future confidence distillation,' a method that trains predictors on pre-solution hidden representations using post-solution correctness probes as teacher signals, enabling reliable confidence estimation before answer generation completes. The paper compares pre-solution Feeling-of-Knowing (FOK) and post-solution Judgement-of-Learning (JOL) confidence estimates across frontier and open-source LLMs, finding post-solution estimates are better calibrated and that linear probes recover richer confidence information than models verbalize. Distilled predictors recover much of the post-solution calibration improvement while remaining sample-efficient and domain-transferable. This has practical implications for confidence-aware systems involving retrieval, tool use, and adaptive computation.
Alignment tuning installs distinct representational directions for sycophancy and cue-induced biases in LLMs
A new arXiv paper investigates where sycophancy and related cue-induced biases live inside LLMs, studying five model families and seven bias types. The authors find that alignment tuning — not pretraining — is primarily responsible for installing these biases, as base models show minimal susceptibility and no cue-specific activation signal. Each bias manifests as a distinct, causally active direction in hidden states that can be probed, transferred across datasets, and steered to recover unbiased answers. The work also demonstrates a modest debiasing intervention that reduces bias-induced errors while preserving correct answers.
DWT-Fusion: Wavelet-based training-free framework for detecting LLM-generated text
Researchers introduce DWT-Fusion, a training-free method for detecting LLM-generated text that applies discrete wavelet transforms to token-level log-probability sequences from a proxy language model, capturing local and multiscale predictability patterns rather than global statistics. The framework evaluates four voting ensemble variants and is tested on HC3, M4, and MAGE benchmarks using GPT-Neo-2.7B, GPT-J-6B, Falcon-7B, and LLaMA-3-8B as proxy models. Best ensemble results achieve AUROC of 0.9919, 0.8477, and 0.7471 on the three benchmarks respectively. The approach is notable for requiring no supervised training while remaining interpretable.
DemoPSD: Disagreement-modulated policy self-distillation to fix privileged information leakage in LLM reasoning training
DemoPSD is a new training framework for LLMs that addresses two failure modes in on-policy self-distillation (OPSD): overfitting to in-domain patterns and privileged information leakage, where the student model learns answer-dependent shortcuts unavailable at test time. The method steers the student toward a reverse-KL barycenter target — a weighted geometric blend of teacher and student distributions — with token-level blending weights derived from the disagreement between the two distributions. Experiments on SciKnowEval across four scientific domains show DemoPSD outperforms GRPO and SDPO while maintaining higher training entropy and generalizing to out-of-distribution GPQA benchmarks.
d-OPSD: First on-policy self-distillation framework tailored for diffusion LLMs
Researchers introduce d-OPSD, the first on-policy self-distillation (OPSD) framework designed specifically for diffusion large language models (dLLMs). The method addresses a fundamental mismatch between existing autoregressive OPSD approaches and dLLMs' arbitrary-order generation by using suffix conditioning on self-generated answers and step-level rather than token-level divergence supervision. Across four reasoning benchmarks, d-OPSD outperforms RLVR and SFT baselines while requiring only ~10% of the optimization steps of RLVR, suggesting strong sample efficiency gains for dLLM post-training.
Self-Policy Distillation via Capability-Selective Subspace Projection
This paper introduces Self-Policy Distillation (SPD), a self-distillation method for LLMs that requires no external signals such as correctness filters or reward models. SPD extracts a low-rank capability subspace from the model's own gradients on correctness-defining tokens, then projects KV activations into this subspace during self-generation to isolate task-relevant signal from stylistic noise. Experiments across code generation, math reasoning, and QA show up to 13% improvement over prior signal-free self-distillation methods and 15% better out-of-domain generalization.