What does this research mean for the field?

A fixed-size, surprise-gated Hebbian associative memory using amplitude encoding significantly improves the perplexity and recall of frozen language models at inference time while using a fraction of the memory required by unbounded kNN-LM approaches. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To evaluate whether a compact, fixed-size Hebbian associative memory with surprise gating and amplitude encoding enables gradient-free inference adaptation in frozen language models.

August 26, 2026Open Access

Sillage: Surprise-Gated Amplitude Memory for Frozen Language Models

View Full Paper

Ask AI

Bookmark

Authors

ASabderrahmane sghairi

Discussion

Member takes

Overview

Computational evaluation demonstrates improved test perplexity in frozen language models via surprise-gated amplitude memory, suggesting efficient online adaptation without backpropagation.

Key Points

To evaluate whether a compact, fixed-size Hebbian associative memory with surprise gating and amplitude encoding enables gradient-free inference adaptation in frozen language models.
Integrated a 4.2 MB associative memory matrix using random token hypervector n-grams, square-root amplitude embeddings, and token surprise gating (-ln p_LM) into frozen GPT-2 and Qwen3-0.6B models.
Evaluated performance across 36k-token and 500k-token text streams against unbounded kNN-LM, byte-matched kNN-LM, exact n-gram cache, and retrieval-augmented rescoring baselines.
Sillage improved GPT-2 test negative log-likelihood by +0.486 ± 0.005 nats (perplexity decreased from 31.2 to 19.2), outperforming unbounded kNN-LM (+0.281 nats, P=1.000) while using 13-fold less memory.
Surprise gating quadrupled the performance gain of uniform writes, while amplitude encoding contributed +0.28 nats over standard count encoding (P=1.000 across all seeds).
On downstream content-token recall, augmented models completed recurring technical terms at up to 2.1-fold the frozen baseline accuracy after a single reading pass.

Cite This Study

abderrahmane sghairi (2026) studied this question.

synapsesocial.com/papers/6a8e9b79451774b83f3b43fe https://doi.org/10.5281/zenodo.22079016

View Full Paper

Ask AI

Bookmark

August 26, 2026Open Access

Sillage: Surprise-Gated Amplitude Memory for Frozen Language Models

View Full Paper

Ask AI

Bookmark

Authors

ASabderrahmane sghairi

Discussion

Member takes

Overview

Computational evaluation demonstrates improved test perplexity in frozen language models via surprise-gated amplitude memory, suggesting efficient online adaptation without backpropagation.

Key Points

To evaluate whether a compact, fixed-size Hebbian associative memory with surprise gating and amplitude encoding enables gradient-free inference adaptation in frozen language models.
Integrated a 4.2 MB associative memory matrix using random token hypervector n-grams, square-root amplitude embeddings, and token surprise gating (-ln p_LM) into frozen GPT-2 and Qwen3-0.6B models.
Evaluated performance across 36k-token and 500k-token text streams against unbounded kNN-LM, byte-matched kNN-LM, exact n-gram cache, and retrieval-augmented rescoring baselines.
Sillage improved GPT-2 test negative log-likelihood by +0.486 ± 0.005 nats (perplexity decreased from 31.2 to 19.2), outperforming unbounded kNN-LM (+0.281 nats, P=1.000) while using 13-fold less memory.
Surprise gating quadrupled the performance gain of uniform writes, while amplitude encoding contributed +0.28 nats over standard count encoding (P=1.000 across all seeds).
On downstream content-token recall, augmented models completed recurring technical terms at up to 2.1-fold the frozen baseline accuracy after a single reading pass.

Cite This Study

abderrahmane sghairi (2026) studied this question.

synapsesocial.com/papers/6a8e9b79451774b83f3b43fe https://doi.org/10.5281/zenodo.22079016

View Full Paper

Ask AI

Bookmark