acceptodds
Under review as a conference paper at ICLR 2027

Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models

Abstract

We present a feedforward graph architecture in which heterogeneous frozen large language models serve as computational nodes, communicating through a shared continuous latent space via learned linear projections. Recent work has demonstrated geometric compatibility between independently trained LLM latent spaces, suggesting that cross-architecture activation translation is possible via linear projection. This work extends that finding from static two-model steering to end-to-end trainable multi-node graphs, where projection matrices are optimized jointly via backpropagation through residual stream injection hooks. Three small frozen models (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) project into a shared latent space injected into two larger frozen models (Phi-3-mini, Mistral-7B), whose representations feed a cross-attention output node. With only 17.6M trainable parameters against approximately 12B frozen, the architecture achieves 87.3% on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on MMLU, outperforming the best single constituent model by 11.4, 6.2, and 1.2 points and outperforming majority vote of the same five models by 8.2, 8.0, and 10.4 points — benchmarks where output-level ensembling actively degrades performance. Gradient flow through frozen model boundaries is empirically verified to be tractable, and the output node develops selective routing toward Phi-3-mini without supervision, validated by a 7.8pp performance collapse upon removal.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
language models, model composition, ensemble methods, latent space, representation learning, frozen models, parameter-efficient methods, multi-agent systems, cross-architecture alignment, residual stream, activation injection, graph neural networks, emergent routing, knowledge aggregation, few-shot learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Mm8AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a feedforward composition of five heterogeneous frozen language models. Three source models produce normalized hidden-state vectors, which are linearly projected and averaged into a shared latent representation. This representation is injected into the residual streams of two downstream frozen models, whose representations are combined by a cross-attention classifier. The projection matrices and output module are trained using answer-level cross-entropy, while the language-model weights remain frozen.

With approximately 17.6 million trainable parameters, the paper reports 87.3% accuracy on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on MMLU, outperforming the reported constituent-model, majority-vote, and single-model learned-head baselines. Its additional contributions are ablations of architectural choices and analyses of gradient flow and component importance.

Critical strengths

1. The reported science-question-answering improvements are substantial enough to merit further investigation.

Sec 4.2 reports improvements of 11.4 percentage points on ARC-Challenge and 6.2 points on OpenBookQA over the best individual constituent. Improvements also remain relative to the reported single-model learned-head baselines. These results make the proposed composition worth studying, even though the current comparisons do not establish which part of the system causes the gains.

2. The ablations provide useful information about the implemented system’s dependencies and failure regime.

Sec 4.4 goes beyond an aggregate accuracy comparison. The injection-coefficient experiment identifies a sharp failure at α=0.75\alpha=0.75, rather than suggesting that stronger intervention is uniformly beneficial. These observations help identify which components deserve further investigation and where the architecture is fragile. They support component dependence, although not all of the proposed explanations for that dependence.

Critical weaknesses

1. The novelty is insufficiently established relative to a directly related prior work.

The paper needs a substantive comparison with LMNet [1], which is a highly relevant but missing related work. LMNet already organizes language-model modules into a layer-wise feedforward graph with trainable dense communication and end-to-end optimization from final-task supervision. Importantly, its Section 4.2.2 evaluates training communication edges while freezing the pretrained vertex models. Its Section 6.1 also explicitly allows different pretrained models at different vertices, although the reported implementations emphasize parameter sharing. Thus, neither dense differentiable model graphs nor frozen-backbone communication learning can be treated as new at the level of the general formulation.

There are architectural differences: LMNet communicates dense sequences through stripped transformer modules, whereas this submission pools final-token representations into a shared vector, performs late residual-stream injection, and uses a fixed four-class output head. The plausible contribution is therefore this specific lightweight cross-architecture implementation and its empirical advantages, rather than the general idea of learning communication among language-model nodes. It should be compared empirically.

2. The experimental design does not yet establish the claimed source of improvement or the reliability of all headline conclusions.

The principal learned control is a classifier on the strongest single model, whereas the proposed system has access to several models and modifies downstream computation. Beating that control does not isolate communication from access to multiple representations. Similarly, comparing a task-trained graph against untrained majority voting does not establish that latent communication is superior to learned output-level combination, where additional training is introduced.

The most important missing control is a jointly trained aggregator over the same five models’ unmodified representations, using the same training data and a comparable trainable-parameter budget. A no-injection variant and a supervised output-combination baseline would further distinguish the value of residual intervention from ordinary learned aggregation. These controls could directly test the paper’s central claim.

[1] Language Model Networks: Supervision-Efficient Learning through Dense Communication. ICML 2026/arxiv: 2505.12741.

Questions for authors

Please refer to Critical Weaknesses.

Reviewer_tycxOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the composition of heterogeneous frozen language models into a trainable feedforward graph. Three smaller models produce normalized hidden states, which learned projections map into a shared latent space and average. This representation is injected into two larger frozen models, whose representations feed a cross-attention classifier. Only the projections and output module, comprising approximately 17.6 million parameters, are trained. The implementation uses a three-to-two topology, shared latent aggregation, and residual-stream injection. The paper reports accuracies of 87.3% on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on its MMLU evaluation, exceeding individual-model, majority-vote, and single-model learned-head baselines. Ablations examine injection strength, latent dimension, node removal, and prompting, while gradient diagnostics are used to interpret model interactions. The evaluation is restricted to supervised multiple-choice classification with one fixed collection of models and single-seed experiments.

Critical strengths

(1) A concrete implementation of frozen-model composition. Section 3 and Figure 1 describe a pipeline for combining heterogeneous internal representations while training only connecting projections and an output module. This provides a useful experimental setting for studying model reuse through continuous communication without updating the constituent backbone weights. (2) Promising reported improvements on multiple-choice benchmarks. Table 1 reports gains of 11.4 and 6.2 percentage points over the strongest constituent model on ARC-Challenge and OpenBookQA, respectively. Improvements also persist against single-model learned-head baselines. These results motivate further investigation, although their interpretation depends on clarification of the evaluation and model-selection protocols. (3) Informative architectural ablations. Table 2 examines relevant design choices, showing sensitivity to excessive injection strength, competitive performance with a smaller latent dimension, and substantial dependence on Phi-3-mini. These experiments help characterize the tested configuration. However, their single-seed setting and differing full-model reference scores across ablation groups limit conclusions about robustness.

Critical weaknesses

(1) The evaluation and model-selection protocols require substantial clarification. Section 4.1 describes 14,042 MMLU examples as training data, matching the official test-split size; the reported 1,531 validation examples also match the official validation split. This suggests a nonstandard supervised protocol, although it does not establish train–evaluation overlap. The authors should identify the exact training, selection, and final evaluation sets and explain how the reported MMLU score should be interpreted. Appendix D also explicitly cites the OpenBookQA test-accuracy plateau as motivation for stopping training, raising concerns about test-informed model selection. Finally, “the same zero-shot greedy decoding protocol” does not adequately distinguish task-supervised graph classification from unadapted constituent-model decoding. Split usage, supervision, answer scoring, and checkpoint selection should be documented, with independent held-out evaluation if test results influenced training decisions. (2) Existing baselines do not isolate the benefit of cross-model injection. The learned-head baseline accesses one frozen model, whereas the proposed graph accesses five models and their internal representations. Its advantage therefore does not establish that residual-stream communication causes the gains, contrary to the attribution in Section 4.2. A central missing baseline is supervised fusion of the same five frozen models without cross-model injection, using the same data and a comparable trainable parameter budget. Disabled or shuffled injection controls would provide complementary evidence, but models trained under these conditions should be distinguished from interventions applied only at evaluation time, which can introduce distribution shift. (3) The contribution relative to prior frozen-model composition is insufficiently established. CALM (Bansal et al., 2024) already composes frozen language models through learned projections and cross-attention, adds the resulting signal through residual connections, and discusses extension to multiple augmenting models. The present contribution therefore centers on its particular multi-node topology, shared mean aggregation, and interpolation-based injection design. Section 2 should explicitly position the method against this closely related work. Comparisons with simpler composition architectures are also needed to establish the value of these specific choices beyond the general principle of trainable communication between frozen models.

Questions for authors
  1. Which exact MMLU splits were used for training, checkpoint selection, and final evaluation, given that the reported 14,042 training examples correspond to the official test split? For OpenBookQA, did test accuracy influence the stopping decision described in Appendix D, or was the training duration fixed independently?
  2. What architecture does “2×L1 only” denote in Table 2, and were the node-removal variants retrained or evaluated by removing nodes from an already trained graph? Were all variants compared under otherwise identical configurations? These details determine whether the results establish the benefit of the graph structure or sensitivity to disrupting a trained system.
Reviewer_jqxjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a fixed architecture of five small language models to do multiple choice question answering benchmarks; they train a small attention layer on the outputs of the networks to improve performance.

Critical strengths
  • Reasonable experimental setup with a variety of ablations.
  • Somewhat novel, in that I have not seen a paper do this exact method.
Critical weaknesses
  • I don't understand why alpha was chosen by hand rather than learned, since the "analysis" section of the paper demonstrates passing gradients backward through entire frozen networks.
  • It's unclear which architecture the phi routing result actually applies to.
  • It's not entirely clear to me how the parameter-matched MLP (or linear layer?) was added to the architecture as a control. If this is just a rotation before the unembedding matrix of the model, I don't think this actually has the potential to improve results very much. Given that the authors actually noticed degradation from this step, I suspect something is wrong here. It would probably be a more fair comparison to add a comparable amount of LoRA parameters to one or more of the models, since they add MHA anyways.
  • Though it seems that the authors did use a train/test/validation split when training their adapters, I would expect small models to benefit directly from training on a task of the same format, regardless of the architecture proposed. That is, I consider this to be overfitting to some extent. A useful baseline would be LoRA in the same setup (the model will adapt to this particular task format).
  • In Section 3.1, pip_i is used without definition (though I suppose it refers to the prefixes in the earlier paragraphs).
  • Linear interpolation to the target dimension seems like a very strange choice. Why not at least use an adapter matrix, given that you are already training other parts of the architecture?
  • A graph structure is mentioned formally but never really used in the paper.
Questions for authors
  • I recommend trying parameter-matched LoRA in the same setting as the one studied in the paper.
  • What exactly is meant by linear interpolation to the target dimension?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.