Back to feed
News Story
APriority83
量子位
1 sources

COLM 2026 Papers Challenge Transformer Defaults, Revealing Hidden Costs in Long-Context and Gradient Flow

At COLM 2026, multiple papers simultaneously question Transformer's default architectural choices, such as QK normalization, GQA, and sliding window attention, finding that these designs underperform in long-context tasks and that combining them exacerbates performance drops. Another study shows that the output projection layer loses most gradient information during backpropagation, hindering training efficiency. These findings may prompt a reevaluation of seemingly reasonable default configurations in future model designs.

SynthePulse Insight · AI deep readingMembers

The 'Standard' Components of Transformers Are Being Collectively Reexamined: Three COLM 2026 Papers Reveal Hidden Costs

Version 1 · 1 source

Long context, gradient flow, layer pruning—those designs that are default in mainstream models and rarely questioned are being reexamined in three independent papers at COLM 2026, which find they exact real costs in evaluation blind spots.

  • Ai2 and others trained 26 7B-8B models, finding that architectural combinations greatly impact long-context performance, with HELMET 32K scores differing by 47%.
  • Cornell University found that the output projection layer erases 95%-99% of gradient norm during backpropagation, and the remaining signal direction is severely distorted.
  • University of Tübingen found that layer pruning causes performance drops on long-chain reasoning tasks, test-time scaling fails, and fine-tuning cannot easily recover.
Open section navigationCracks in Long Context: The Hidden Cost of Architectural Combinations

Cracks in Long Context: The Hidden Cost of Architectural Combinations

At COLM 2026, the paper 'Cracks in the Foundation' from Ai2 and others systematically challenges several standard Transformer designs: QK normalization, GQA, sliding window attention, and pretraining context length. The researchers selected different values from Llama 2, Llama 3, Qwen 3, and Olmo 3, training 26 models with 7B-8B parameters, collectively called 'OlmPool', keeping data, tokenizer, and context extension scheme consistent, only varying architecture.

These models were first pretrained on 140B tokens, then fine-tuned for long context with 10B tokens, consuming over 170,000 GPU hours in total. On the HELMET 32K benchmark, the highest score was 56.4 and the lowest 29.9, a relative gap of 47%. Adjusting a single switch (such as pretraining context length or sliding window) on average affects only one or two points, but combined effects are significant: SWA and GQA together drop an average of 9 points, exceeding the sum of their individual effects; the worst combination (GQA+SWA+per-head QK normalization) also drops more than linearly.

A counterintuitive finding is that QK normalization (originally added for training stability) when removed and replaced with pre-normalization on Olmo 3 improves HELMET scores by 6 points, but the same change on Llama 3 decreases by 3.8 points, indicating that the same design can have opposite effects on different models. Additionally, the paper finds that 'attention sinking' (where models concentrate a large amount of attention on tokens at the beginning of the context) is positively correlated with long-context performance in OlmoPool models, and QK normalization weakens this sinking, which may explain its negative impact.

Free for now

Read the full analysis

3 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article is based on QbitAI's report on three COLM 2026 papers, and all data comes from that report. The papers themselves are preprints or conference papers, not peer-reviewed, and the report does not provide the full text of the original papers, so some conclusions (such as the specific mechanism of gradient loss) are claims by the paper authors, not independently verified facts.

Primary report

量子位

Primary source