acceptodds
Under review as a conference paper at ICLR 2027

Global Latent Workspace: Shaping the Workspace for Multi-Agent Latent Communication via Reinforced Attention

27 views

Abstract

Latent communication allows multi-agent LLMs to exchange internal representations without repeatedly decoding and re-encoding text. However, relaying intermediate states without selective filtering can introduce redundancy and tie communication cost to local reasoning length. We propose the Global Latent Workspace (GLW), a bounded communication interface that learns which information to share for downstream reasoning. Independent experts generate local latent states, and a lightweight controller selectively integrates them into a fixed-size workspace. After workspace initialization, the controller skips or blends each remaining candidate. A frozen final agent reads the resulting workspace to produce the answer. This decouples the amount of local computation from shared-context size, making communication a learnable component of multi-agent reasoning. We train only the controller using reinforcement learning with attention-derived feedback and writer-level counterfactual rewards, while keeping all agent backbones frozen. Experiments on six benchmarks with two LLM backbones show strong reasoning performance relative to three representative baselines. GLW demonstrates strong reasoning performance across all six tasks with the 8B backbone, while reducing shared-context length relative to latent relay and generated output relative to text-based collaboration. A fixed-budget state-selection study examines learned admission; reward ablations assess the contributions of frozen-judger feedback.

Primary area:
foundation or frontier models, including LLMs
Keywords:
latent communication, agentic systems, MAS

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8Xh7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Global Latent Workspace (GLW), which is a learned communication interface for a particular set of agent workflows. Three independently prompted experts first generate 120 continuous candidate states, 40 states initialize a workspace, and a controller skips or blends the remaining candidates into those positions. Then a final frozen reader consumes the workspace as embeddings. The controller evaluates the candidate vectors and decide when to skip / rewrite the workspace vectors that are used as context for the reader model, and it is trained to maximize final outcome correctness; Everything else is frozen.

The evaluation covers 6 multiple choice benchmarks with Qwen3-4B/8B. Training is done on ARC-Challenge and transferability is observed on other datasets (QASC/SciQ/CommonsenseQA/MedMCQA). GLW exceeds LatentMAS baseline on all 6 8B tasks and 5 out of 6 4B tasks.

Critical strengths
  1. The system setup is clear and easy to understand.

  2. The benchmark performance improvement over LatentMAS is meaningful, especially when it seems to have transferred to other datasets.

Critical weaknesses
  1. The improvement over a TextMAS is less substantial when compared to LatentMAS (0.33 - 2.67 pp across 8B, and regressions on w/6 tasks at 4B). Although the generated text is substantially reduced, the reported communication metrics do not establish whether its overall accuracy-cost tradeoff justifies controller training.

  2. The main transfer results use ARC-Challenge as the controller-training source and evaluated a limited collection of multiple-choice tasks. Although CommonsenseQA and MedMCQA provide useful shifts, the results leave open how transfer depends on the training source and how well the learned communication extends to more distinct domains.

Questions for authors
  1. To address weakness 1, it would be beneficial to report the controller-training cost and complete inference cost on specific hardware, alongside accuracy uncertainty (e.g. paired confidence interval) for GLW and TextMAS?

  2. To address weakness 2, it would be beneficial to a) include questions from a more distinct domain such as LegalMCQA, and b) train another controller on, e.g. CommonsenseQA, and evaluate the controllers on the same else datasets - ARC, MedMCQA, SciQ etc.

  3. Clarification: could you specify if you choose one controller version using ARC validation and reuse it everywhere, or use each benchmark’s validation results to choose a version?

Reviewer_ZksZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a Global Latent Workspace (GLW) for communication among frozen LLM agents. Three independent, role-prompted experts each generate 40 aligned latent vectors. The main deployment initializes a 40-position workspace from the first interleaved candidates and processes the remaining 80 with learned skip/blend decisions and content-addressed erase/add updates. A frozen judger re-encodes the workspace and answers the question. Controller training combines writer-suppressed outcome contrasts, attention-derived rewards, a grouped policy objective, and a reference-based content loss. Experiments cover six multiple-choice benchmarks with Qwen3-4B/8B and separate selection/reward controls.

Critical strengths
  1. The fixed workspace gives a useful accuracy/context-size tradeoff. Expert computation is separated from the capacity available to the final reader. The comparison with LatentMAS uses matched expert prompts and latent-step budgets, which makes this tradeoff easier to assess. On 8B, GLW exceeds all three baselines on all six tasks; ARC-Challenge reaches 97.17%, versus 90.50% for LatentMAS and 94.50% for TextMAS. These results make the workspace worth studying, even though the source of the gains needs further clarification.

  2. The ablations cover several plausible alternative explanations. Random, fixed-first, untrained, and zero-vector selection controls test whether learned selection matters, while uniform/shuffled/disabled attention rewards examine the reward design. The manuscript acknowledges that the reported reward gains are small and distinguishes communication metrics from total computation. This makes it easier to judge which improvements the experiments actually support.

Critical weaknesses
  1. The proposed continuous-update mechanism is not isolated by the strongest controls. Table 1 uses sequential skip/blend updates; Appendix B.1 instead gathers 40 original vectors with a different selection objective based on option likelihood and correctness. Those results validate learned admission but do not isolate the deployed controller or blending. The selected-vector result exceeds the zero-vector workspace by only 1.67 points. Furthermore, Equation 9 pulls positive-advantage updates toward a reference output and pushes negative-advantage updates away. Reference distance does not specify a direction of downstream improvement; its gradient vanishes when outputs coincide, and negative terms favor increasing distance without an explicit content-space bound. This is not a proof of optimizer divergence, but direct validation of the continuous-write surrogate is missing. The 0.67- and 0.50-point reward gains do not settle this central attribution issue.

  2. The evaluation does not yet establish robust transfer or an end-to-end advantage. Each final evaluation has 200 questions. MedMCQA uses the same slice for validation and final evaluation, so its independence depends on whether that slice influenced selection. Repeating the same questions across seeds does not quantify question-sampling uncertainty, relevant to gains such as 0.17 points on SciQ-8B. CommonsenseQA-4B degrades against every baseline. The context reduction also includes dropping expert prompt KV and switching to fresh encoding, while GLW still computes all candidates and produces more output tokens than LatentMAS on five of six 8B tasks. Without matched interface controls and representative timing, the mechanism's practical benefit remains uncertain.

Questions for authors
  1. Can the deployed skip/blend controller be compared with its initialization and a replacement-only or frozen-write-head variant, using the same candidate bank and 40-position capacity? I would also like to understand what evidence supports Equation 9 as a content-update objective, particularly after reference synchronization.
  2. Which dataset is used to select each transferred checkpoint? Was MedMCQA validation[0:200] used in any checkpoint or configuration selection? If so, can results on a disjoint evaluation slice be provided?
  3. How does the 40-position workspace compare in accuracy and end-to-end cost with a baseline that re-encodes all 120 candidate vectors and also omits expert prompt KV? Such a comparison would clarify how much of the benefit comes from workspace compression.
Reviewer_sL6COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GLW learns a bounded communication interface between frozen, independently prompted LLM experts and a frozen final reader. Three experts generate 120 latent candidates; 40 initialize a workspace, and a controller skips or blends the remaining candidates. Controller training combines writer-suppressed outcome feedback with attention-derived rewards. Evaluation covers six multiple-choice benchmarks and two backbone sizes, supplemented by a separate fixed-budget selection study and reward ablations for the update controller.

Critical strengths
  1. The bounded workspace addresses a concrete communication bottleneck while preserving the local latent-computation budget. On QASC with the 8B backbone, accuracy increases from 85.50% for LatentMAS to 87.50%, while final-reader context decreases from 721.5 to 197.1 positions (Table 1). Keeping the backbones frozen makes this a useful contribution at the communication-interface level.
  2. The evaluation includes informative controls beyond aggregate scores. Fixed-capacity selection comparisons examine learning and candidate content, while observed, uniform, shuffled, and disabled attention conditions examine the feedback signal. The paper appropriately distinguishes these studies and treats the small reward gains as descriptive rather than claiming statistical significance.
Critical weaknesses
  1. The benefit of learning in the deployed skip/blend controller is not fully isolated. The untrained, first-40, random-40, and zero-vector comparisons use a separate gather-40 selector trained with option-likelihood feedback (Appendix B.1). These results do not directly quantify the contribution of the main update controller or its erase/add heads. The reward ablations help, but a matched trained/untrained update-controller comparison remains important.
  2. Validation independence and precision limit part of the empirical evidence. Table A4 lists the same MedMCQA slice for validation and final evaluation; whether it influenced checkpoint selection needs clarification. Each task uses 200 evaluation questions, and repeated seeds do not replace question-level uncertainty estimates. The disclosed limitations constrain interpretation of the smaller gains.
Questions for authors
  1. Can trained and untrained versions of the actual skip/blend controller be compared under identical candidate banks and decoding, with a frozen-write-head or replacement-only control to isolate blending?
  2. Which data selected the MedMCQA checkpoint and configuration? If validation[0:200] informed selection, can you report performance on a disjoint labeled evaluation slice?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.