Near-Zero Token Embeddings Silently Stall Training in Gated Sequence Models
The paper reports that among 291 surveyed open-weight language model repositories, 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below 10^-6, versus 0% of weight-tied models.
Source-based brief · Editor review not recorded · See checks and limits ↓
One primary source; other links are context, not independent corroboration.
- Primary sourceHugging Face
- 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
The paper reports that among 291 surveyed open-weight language model repositories, 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below 10^-6, versus 0% of weight-tied models. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), near-zero tokens can amplify input-embedding gradients sharply, and global-norm clipping can overflow to infinity and zero a batch's gradients. The authors propose GroupClip for released checkpoints and GateFloor for pretraining.
Claim explorer
Attribution, source and conditionsSource-stated points from the stored brief. Claim-level assessments have not been recorded.
Source claims not independently checked; the full paper was reviewed from the fetched page text, and experiments are reported for models up to 370M parameters plus one 4B continual-pretraining run.
This connects a source, an attributed statement and its qualifications. It does not establish that the claim is correct.
Questions worth asking
Open questions for further reportingWhat remains unconfirmed?
Source claims not independently checked; the full paper was reviewed from the fetched page text, and experiments are reported for models up to 370M parameters plus one 4B continual-pretraining run.
What material was available for this brief?
The recorded brief basis is fetched source page. These points are AINewsIQ paraphrases, not quotations or independent confirmations.
Among 38 untied gated and hybrid checkpoints, 20 contain reachable near-zero input embeddings (norm below 10^-6).
Source: Hugging Face ↗Stored source-based brief · key point 1 · AINewsIQ paraphraseIn gated models, a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude.
Source: Hugging Face ↗Stored source-based brief · key point 2 · AINewsIQ paraphraseIn Telugu continual pretraining of Nemotron-3-Nano-4B, the failure discards updates on about 20% of steps.
Source: Hugging Face ↗Stored source-based brief · key point 3 · AINewsIQ paraphraseGateFloor reduces near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with comparable perplexity.
Source: Hugging Face ↗Stored source-based brief · key point 4 · AINewsIQ paraphraseLimits from the stored brief
Source claims not independently checked; the full paper was reviewed from the fetched page text, and experiments are reported for models up to 370M parameters plus one 4B continual-pretraining run.
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.
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.
Brief history · 1 retained version
These are changes to AINewsIQ’s own publication. Earlier history is unavailable unless recorded here. Complete retained versions are available to the editor.
- Version 1 · published
Source-based brief stored or updated. This does not establish that the publisher changed its page.
Automated record
Source observations · 1 recent captures
A changed capture means the extracted text differed. It can reflect page furniture or extraction changes; it does not establish a substantive correction. Source captures remain private.
Retained capture; no earlier comparison available.
New engagement snapshots do not mean the source itself changed. Source observations and AINewsIQ brief revisions are separate records.
Why this may matter
Silent gradient blowups and discarded updates can stall training without obvious errors, affecting practitioners continually pretraining gated or hybrid language models. The paper also offers concrete mitigations.
Limits and open questions
Source claims not independently checked; the full paper was reviewed from the fetched page text, and experiments are reported for models up to 370M parameters plus one 4B continual-pretraining run.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Join the discussion
Have a different perspective or a question about the evidence? Discussion is free with a verified email account. Every comment is reviewed before publication.
No published comments yet.
New commenters: 2 submissions per 24 hours, at least 5 minutes apart, until 3 are approved. Discussion guidelines and moderation →