mlsanityData to 7 Oct 2026

Stuck on "A": Diagnosing and Repairing Interface Injury in Attention-to-KDA Linearization of a 0.6B Language Model

arXiv 2608.02689 ↗First posted 3 Aug 2026cs.CL cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

We convert 21 of 28 full-attention layers of Qwen3-0.6B-Base into KDA (Kimi Delta Attention) linear-attention layers on a single consumer-grade GPU budget, and ask a simple question: what exactly does the conversion break? After surgery, hidden-state alignment and end-to-end KL distillation drive the student close to its teacher in perplexity, yet multiple-choice accuracy stays near random chance (25-29% vs. the teacher's 50.6% on C-Eval). Using a four-permutation diagnostic that rotates answer options while holding content fixed, we show the model sticks to option labels (predicting "A" 81% of the time; 106/161 questions keep the same label under all four rotations) rather than following answer content -- an interface injury that standard distillation metrics cannot see. A 1,000-step format-targeted completion-only KL stage repairs the interface (+12.48 points on C-Eval, label-stickiness roughly halved), after which persona SFT and one round of on-policy DPO preserve benchmark scores within noise. We release code, weights, recipes, and the full audit trail, and distill the engineering lessons -- including an FP32-master failure mode in which bf16 optimizer updates are silently swallowed -- that made convergence possible at this budget.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 8 of its 11 references that are in the map. First posted on arXiv 3 Aug 2026.

Builds on

8 of 11 references.

References · 8

  1. 2015Distilling the Knowledge in a Neural Networkcited by 10,076
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  3. 2020Transformers are RNNs: Fast Autoregressive Transformers with Linear Attentioncited by 1,979
  4. 2023C-Eval: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Modelscited by 694
  5. 2024Gated Delta Networks: Improving Mamba2 with Delta Rulecited by 483
  6. 2025Kimi Linear: An Expressive, Efficient Attention Architecturecited by 169
  7. 2026Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contextscited by 16
  8. 2026When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Modelscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Text Capability Loss in Vision-Language Adaptation: An Attention-Sink Diagnosis
  2. 2026Gradient-Stable Attention Heads Signal LLM Correctness
  3. 2026Linear Predictability of Attention Heads in Large Language Models
  4. 2026What Attention Recalls and Recurrence Controls in Hybrid Language Models
  5. 2026Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits
  6. 2026Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States
  7. 2026Where Should LoRA Go? Component-Type Placement in Hybrid Language Models
  8. 2024Simple linear attention language models balance the recall-throughput tradeoff