ThinkBit: Search Codewords, Recalibrate Scales, Select by Reasoning
ThinkBit is a scalar two-bit quantization framework aimed at preserving multi-step reasoning in language models. It assigns each layer a ternary, 2-bit, or 3-bit tier, searches codewords with a Gumbel-softmax method, recalibrates scales via knowledge distillation, and applies short quantization-aware training. On Qwen3-1.7B, the recalibrated checkpoint scores 46.7 on MATH-500 versus 30.0 for EfficientQAT, and a 14B model at 3.74 GB scores 94.8 against 97.5 for bfloat16.
One primary source; other links are context, not independent corroboration.
- Primary sourceHugging Face Daily Papers
- Discovery linkHugging Face
- Linked in sourceView arXiv page
- Linked in sourceView PDF
The paper page is linked. Peer review and findings have not been independently checked by AINewsIQ.
Open full article view →- Original paper+45
- Fetched source page+30
- No editor link review+0
- Original non-discussion link+5
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
ThinkBit is a scalar two-bit quantization framework aimed at preserving multi-step reasoning in language models. It assigns each layer a ternary, 2-bit, or 3-bit tier, searches codewords with a Gumbel-softmax method, recalibrates scales via knowledge distillation, and applies short quantization-aware training. On Qwen3-1.7B, the recalibrated checkpoint scores 46.7 on MATH-500 versus 30.0 for EfficientQAT, and a 14B model at 3.74 GB scores 94.8 against 97.5 for bfloat16.
Key findings from the stored brief
- On Qwen3-1.7B at matched size, the recalibrated checkpoint scores 46.7 on MATH-500 versus 30.0 for EfficientQAT.
- At 0.52 GB, the Qwen3-1.7B checkpoint is smaller than every published 2-bit GGUF of the model yet more accurate on MATH-500.
- A 14B model at 3.74 GB scores 94.8 on MATH-500, compared with 97.5 for the 29.5 GB bfloat16 model.
Source claims not independently checked; results are reported on MATH-500 with specific models and a paired reasoning test on training problems.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
Related archive coverage
These relationships record a citation or topic overlap, not confirmation of the finding.
The paper page is linked. Peer review and findings have not been independently checked by AINewsIQ.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
A publication revision history is not stored for this signal. New engagement snapshots do not mean the source itself changed.
Why this may matter
Two-bit weights could substantially cut memory bandwidth costs of long reasoning generations. The paper argues that perplexity and multiple-choice accuracy miss reasoning degradation, so evaluating on hard math problems offers a more relevant test for compressed models.
Limits and open questions
Source claims not independently checked; results are reported on MATH-500 with specific models and a paired reasoning test on training problems.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Discussion
Ask about the source, add context, or point readers to a relevant follow-up.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.