acceptodds
Withdrawn from ICLR 2027

Memory Without Training: A Complementary Learning System on Frozen Multimodal Encoders

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We report a complementary learning system that is never trained, and whose every behaviour is derived before it is measured. A fast store binds role–filler pairs by circular convolution and completes them through an attractor read; a slow store accumulates the same content in a Hebbian outer-product matrix under passive decay; replay of confidence-gated retrievals carries episodes from one to the other. Binding capacity, the forgetting horizon, the consolidation schedule, the onset of semanticisation and the magnitude of a priming effect are each predicted from a single interference calculation, with no fitted parameter. Content comes from frozen encoders: MiniLM and CLIP codes enter through a codec whose crosstalk we predict from the codes' power spectra alone, and an untrained random projection into a shared workspace binds language and vision slots into one episode, within 0.011 of the derived accuracy and with no new constant. Three findings stand out. Binding capacity belongs to the workspace dimension rather than to the encoder that produced the content, so projection alone buys 15.3 points of accuracy at depth. The spacing effect reverses direction with the clock it is measured from, and both directions follow from the same law. And the full circuit adds nothing over the storeless codec law, agreeing within 0.009 at every load, so end-to-end behaviour is the product of independently derived parts. We registered all 23 predictions before measuring them; four failed, and we report the failures, the diagnoses they carried, and three corrected laws re-validated on grids they had never seen.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
episodic memory, vector-symbolic architectures, complementary learning systems, multimodal binding, training-free, pre-registration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_jUo2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposed a training-free memory mechanism. The system design cleverly extracts role-filler information while deriving predictable behavior.

Critical strengths

The paper proposes a memory mechanism, showing that memory can be constructed from frozen pre-trained encoders, in a predictable way without any training. This approach may support long-term episodic implications. The authors transparently discuss their results, highlight their limitations, and analyzed their failures.

Critical weaknesses

The paper's contribution is limited. The proposed memory system is complex and feels hand designed to achieve desired behavior, without providing a reason or practical implications of using it beyond no training.

The writing can be improved, for example: experiments are not sufficiently detailed in the main text, the claims in the introduction are provided without proper citation, and the main contribution is not clear enough.

Questions for authors

What benefit does this system propose beyond the lack of training? Can it be used for practical implications on real scenarios or provide new insight on current memory systems?

Reviewer_RbxsOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a training-free multimodal memory system that combines circular-convolution binding, Hebbian storage, confidence-gated replay, and a shared workspace on frozen MiniLM and CLIP encoders. The motivation is reasonable, and the method has some theoretical interpretability, but the experiments do not sufficiently demonstrate its practical advantages or generalization ability.

Critical strengths
  • It integrates VSA, Hebbian storage, and complementary learning into a relatively complete framework.
  • It uses preregistered experiments and reports failed predictions, providing some research transparency.
Critical weaknesses
  • The manuscript is poorly organized. The connections among the background, motivation, and contributions are unclear, making it difficult to read.
  • The figures are of low quality. Their fonts, legends, axis labels, and visual encodings are not sufficiently clear.
  • The experimental results do not sufficiently demonstrate the advantages of the proposed method. In addition, the datasets, baselines, and experimental settings are insufficiently described, making it difficult to assess the fairness of the comparisons and the credibility of the results.
Questions for authors

See Critical Weaknesses.

Reviewer_zPABOpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a memory system built on frozen MiniLM and CLIP encoders, without additional gradient-based training. A fast episodic store binds role–filler pairs through circular convolution and permutations and retrieves episodes using an attractor mechanism. Confidence-gated replay transfers retrieved fillers into a decaying Hebbian store using event-specific index keys. Random projections map different encoder outputs into a common workspace. The paper’s main claim is that memory capacity, forgetting, and replay effects can be predicted quantitatively from interference calculations, which it evaluates through controlled simulations.

Critical strengths

The paper preserves failed claims rather than quietly replacing them. Forgetting and retention discrepancies motivate revised predictions evaluated on fresh parameter grids. The supplied registry, lock, source fingerprints, and caches are internally consistent. This is useful evidence of disciplined bookkeeping, although it does not establish when predictions were made.

Critical weaknesses
  • This paper is a bit hard to read and the writing clearly needs improvements. The introduction and abstract reads more like prose instead of a technical report.
  • The references are not linked at all (e.g., the section reference in the introduction, the bib text in the related work).
  • The experiments query the slow memory using the exact event-specific keys retained during writing. The paper does not explain how a later content-based query obtains these keys, particularly after the fast memory is removed. Consequently, the experiments demonstrate retention given a known address, but leave end-to-end retrieval from consolidated memory unestablished.
Questions for authors
  • Can you give a worked end-to-end example and summarize the numerical predictors in the paper?
  • Can the missing embeddings/text lists be regenerated from the supplied package, and can the preregistration chronology be audited?
Reviewer_y6itOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a training-free complementary learning system with a fast store and a slow store, and a confidence-gated replay between them. This system works as a multimodal memory without requiring neural network training. The authors derive the memory behavior mathematically in advance using a single interference calculation.

Critical strengths
  1. The paper reports failed predictions, along with diagnoses, showing scientific integrity.
  2. The paper derives complex system behaviour from parameter-free equations rather than relying on learned weights.
Critical weaknesses
  1. The paper's presentation is really not clear. In the introduction, not a single related work is mentioned. It is not clear from the introduction what the research question is, what has been done, and what they do and why it is important.
  2. The experimental setup is very abstract. There's no performance on standard benchmark is reported. It makes the utility of this memory system questionable
  3. Experimental details are missing, and readers cannot reconstruct the experiments from it.
Questions for authors

See weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.