The paper introduces Selective Rank-Preserving Supervised Fine‑Tuning (SRP‑SFT), a method that applies a hinge‑based soft margin target to preserve selected local non‑target logit margins during teacher‑forced fine‑tuning. SRP‑SFT uses reference‑scaled lower bounds for pairs from the reference’s top‑K non‑target tokens, stopping once the bound is met. Experiments on Qwen3‑4B‑Base and UltraFeedback show that SRP‑SFT improves MATH‑500 and code‑generation scores compared to cross‑entropy and LP‑SFT baselines.
arXiv:2607. 22334v1 Announce Type: new Abstract: Open-weight language models from different families exhibit complementary capabilities, motivating their consolidation into a compact student through on-policy distillation (OPD).
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:2607. 12204v1 Announce Type: new Abstract: Attention can be viewed as an online learner over context, yet existing test-time memories cannot certify that dropping a token leaves outputs unchanged or delete its influence outright.
The paper introduces PACT, a tuning method for single-token typed-decision models that leverages contrastive pair data to add four training terms—difference-in-differences margin, permutation-consistency, evidence-necessity, and ordinal transport cost—without requiring new annotations. PACT achieves comparable accuracy to existing recipes while reducing position bias and ordinal error, and it improves robustness and stability across seeds. The authors provide code, data splits, and trained adapters for reproducibility.
The paper demonstrates that a byte‑level BPE tokenizer can be sliced to create multiple vocabulary sizes from a single trained model, preserving exact logits while reducing deployed weights by 66%. Experiments on 30 models show that while sliced models match the full model numerically, they underperform fixed‑cap specialists by a few percentage points in bits‑per‑byte. Multi‑cap training improves robustness to typographical noise, suggesting benefits from training across multiple granularities rather than from control tokens alone.
arXiv:2607. 15498v1 Announce Type: cross Abstract: The key-value (KV) cache is the main memory bottleneck in long-context large language model (LLM) inference.
arXiv:2609.39687v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises studen...
TwinKV is a training‑free, attention‑free repair pass that identifies and swaps orphaned and redundant tokens in a KV cache, improving long‑context inference for small models. It works by detecting near‑duplicate keys and can be composed with existing eviction policies without altering their scoring rules. Experiments on Qwen3‑4B and Llama‑3.2‑1B across LongBench, LooGLE, RULER, and MMLU‑Pro show that TwinKV consistently improves performance for most configurations, especially at tighter compression ratios.
ValueDiff introduces a value‑geometric KV cache eviction strategy for large language models that suppress attention sinks. It ranks tokens by the L2 deviation of their value vectors from the cache mean, a score that aligns with minimal‑disturbance eviction under a max‑entropy assumption. Across several benchmarks—RULER, LongBench, and MATH‑500—ValueDiff consistently retains a higher proportion of useful tokens than prior methods, especially under tight cache budgets.
arXiv:2611. 07232v1 Announce Type: cross Abstract: At context lengths of about 1M tokens, dense attention is costly in prefill and decoding, and training typically relies on long documents, which become scarcer as the target length increases.
GEAR is a two‑stage framework that distills tabular foundation models into lightweight MLP or tree‑based predictors for efficient CPU deployment. In the first stage, synthetic covariates are used as teacher‑query locations to train the student on soft TFM targets, expanding coverage beyond observed rows. The second stage re‑anchors the student to the target distribution using real labels and out‑of‑fold teacher predictions, preventing self‑labeling leakage and improving performance. Experiments on TALENT and TabArena show that GEAR‑distilled MLPs outperform supervised MLPs by up to 2.00 AUC points on binary tasks and 1.35 on multiclass tasks, and also outperform CatBoost, while dramatically reducing inference time and memory usage.
arXiv:2609.30738v1 Announce Type: new Abstract: KV cache eviction methods such as SnapKV and PyramidKV rank tokens solely by mean attention over a small observation window. We study a unified score,...