XTC (Exclude Top Choices) is a lightweight, head‑aware decoding operator that improves diversity in autoregressive language models by removing overly probable tokens that dominate the next‑token distribution. It works by identifying tokens above a plausibility threshold, probabilistically excluding the dominant choices, and renormalizing the remaining distribution. Across 60 experiments on models such as Gemma 3 and DeepSeek R1, XTC boosts Distinct‑2 scores by 11–15 % and cuts repeat trigrams by 27–47 %, while a Mechanical Turk study shows a 62.3 % preference for XTC‑generated text without loss of fluency.
Standard decoding rules for autoregressive language models promote diversity by rescaling the full next-token distribution or truncating its low-probability tail. These strategies overlook a common re...
arXiv:2609.07901v1 Announce Type: new Abstract: Weight quantization largely determines the economics of serving open-weight LLMs. Its costs are usually assessed with capability benchmarks, on which 4...
arXiv:2410.02343v2 Announce Type: replace Abstract: Large language models (LLMs) routinely fail to output the correct option in multiple-choice question answering (MCQA) while encoding the answer int...
arXiv:2606. 01561v1 Announce Type: new Abstract: Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO).
arXiv:2610.19996v1 Announce Type: new Abstract: Large language models often produce fluent but factually incorrect output. Gray-box detectors identify such hallucinations from the token-probability s...
The paper introduces Variance‑Calibrated Modulation (VCM), a training‑free pre‑decoding technique that reshapes language model probability distributions before truncation. VCM uses two dynamic mechanisms: a Contextual Searchlight via PMI to suppress stopwords and highlight context‑relevant tokens, and an Adaptive Self‑Debiasing that applies scale‑invariant penalization based on real‑time logit standard deviation. Experiments on open‑ended generation, factual QA, and mathematical reasoning show that VCM consistently reduces the likelihood trap, improving diversity, coherence, and reasoning accuracy with minimal computational cost.
Almost every large language model that reaches a broad audience is quantized: trained in full precision, then compressed for efficiency. This step is assumed harmless and its safety is rarely re-checked.
The paper investigates how preference optimization can inadvertently create a shared directional shift—termed a common‑mode offset—in a model’s hidden states. This offset, which affects both chosen and rejected responses, accumulates without being penalized by the ranking loss, eventually projecting onto the output layer to form a vocabulary well that dominates decoding and leads to generation collapse. Experiments across multiple base‑model lineages and fine‑tuning regimes confirm that masking this offset or using per‑token ranking losses can prevent the collapse, while standard ranking metrics fail to detect the risk.
arXiv:2510. 01171v4 Announce Type: replace-cross Abstract: Post-training alignment often reduces LLM diversity, leading to a phenomenon known as mode collapse.
arXiv:2607.27836v2 Announce Type: replace Abstract: Large language model unlearning is consistently fragile under relearn attacks. On TOFU, fine-tuning on twenty forget examples substantially recover...
arXiv:2609.14896v1 Announce Type: cross Abstract: A notable byproduct of LLM alignment training is mode collapse: the progressive loss of output diversity that narrows a model's expressivity at infer...
arXiv:2606. 00544v1 Announce Type: new Abstract: Modern language-model fine-tuning typically pairs each prompt with a single response, even though many prompts admit multiple valid completions.