Research map / Large language models / Language, safety and interpretability
Sudden Drops in the Loss: Syntax Acquisition, Phase Transitions, and Simplicity Bias in MLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies how syntactic attention structure (SAS) emerges abruptly during BERT-base MLM pretraining, shows via a syntactic attention regularizer that SAS is necessary for grammatical capabilities, and finds briefly suppressing SAS improves model quality.
Problem
Understanding how interpretable internal structure (syntactic attention heads) emerges during MLM training and whether it causally drives grammatical capabilities, which test-time analyses cannot establish.
Approach
Developmental analysis of training trajectories combined with a syntactic attention regularizer that promotes or suppresses SAS during pretraining, revealing a structure onset that precipitates a capabilities onset, a competing alternative strategy, and benefits of early brief SAS suppression.
Key results
- BERT-SAS- (regularizer suppressed throughout) shows a loss break earlier (~6K steps), an alternative strategy onset, while BERT-SAS+ reaches MLM loss 2.39 vs 1.77 for BERT-Base at 100K steps.
- The multistage model BERT-SAS-(3k), with SAS suppressed until step 3000, reaches MLM loss 1.75 vs 1.77 for BERT-Base and BLiMP 0.75 vs 0.74, with GLUE 0.74 vs 0.74 at 100K steps.
- Releasing the SAS regularizer during the alternative strategy onset yields the worst model quality across metrics, suggesting breakthroughs are critical points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| promote SAS throughout training | BERT Base · 100K training steps | GLUE average: 0.74 \pm 0.01 -> 0.59 \pm 0.01 | Hurt |
| promote SAS throughout training | BERT Base · 100K training steps | BLiMP average: 0.74 \pm 0.02 -> 0.74 \pm 0.01 | No clear effect |
| suppress SAS throughout training | BERT Base · 100K training steps | BLiMP average: 0.74 \pm 0.02 -> 0.67 \pm 0.03 | Hurt |
| suppress SAS until 3000 steps, then release the regularizer (ablation) | BERT Base · 100K training steps | MLM loss: 1.77 \pm 0.01 -> 1.75 \pm 0.01 | Helped |
| suppress SAS until 3000 steps, then release the regularizer (ablation) | BERT Base · 100K training steps | BLiMP average: 0.74 \pm 0.02 -> 0.75 \pm 0.01 | Helped |
| suppress SAS only briefly early in training (ablation) | BERT Base · training trajectory | structure onset timing | Helped |
| continue suppressing SAS past the alternative strategy onset (ablation) | BERT Base · training trajectory | model quality | Hurt |
| change the regularizer during the alternative strategy transition (ablation) | BERT Base · 100K training steps | model quality | Hurt |
| suppress SAS throughout training | BERT Base · BLiMP grammatical capabilities | BLiMP accuracy | Hurt |
Surprising findings
- Briefly suppressing syntactic attention structure early in training improves final MLM loss and BLiMP over the baseline.
- Releasing the regularizer during a phase transition produces a worse model than releasing it before or after.
Limitations
- Correlational analysis across MultiBERTs seeds found no significant correlation between UAS and capabilities, so SAS necessity relies on interventions.
- Results are for BERT-Base with a single architecture and a small pretraining corpus.
- Silver-standard spaCy parses are used for the syntactic regularizer and probing.
Experimental setup
Datasets: BookCorpus English Wikipedia (May 2020 dump) Penn Treebank (WSJ, 1000 documents) GLUE BLiMP
Benchmarks: BLiMP GLUE Penn Treebank WSJ UAS
Metrics: MLM loss UAS BLiMP accuracy GLUE average
Baselines: BERT-Base BERT-SAS+ BERT-SAS-
Builds on: BERT simplicity bias Clark et al. attention-head syntax probing
Closest papers by meaning
- Understanding Emergent Abilities of Language Models from the Loss Perspective (NeurIPS 2024 · similarity 0.88)
- Training on the Test Task Confounds Evaluation and Emergence (ICLR 2025 · similarity 0.87)
- Emergence of Hidden Capabilities: Exploring Learning Dynamics in Concept Space (NeurIPS 2024 · similarity 0.87)
- An Analysis for Reasoning Bias of Language Models with Small Initialization (ICML 2025 · similarity 0.87)
- Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale (NeurIPS 2025 · similarity 0.87)
- Interpreting the Repeated Token Phenomenon in Large Language Models (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.