{"id":"6377cbee-3706-41bd-ba07-ad0bea68b5ca","arxiv_id":"2608.13317","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"StateBridge aligns sender final-layer hidden states to the receiver's embedding space via closed-form orthogonal Procrustes plus norm and vocabulary calibration, enabling training-free latent communication that outperforms text and KV-cache baselines.","lead":"A new method lets LLM agents exchange hidden neural states instead of text, using a training-free rotation and calibration step. It beats text and prior latent baselines on most of 26 model-task pairs, with the largest gains on harder reasoning and code tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Only anecdotal evidence supports the load-bearing claim that the final-K-state prefix encodes more than the last K tokens; a suffix-text control baseline would settle it.","rationale":"StateBridge is clearly presented and the closed-form alignment math is correct; the ablations and the K/alpha sensitivity study are useful. However, the scientific novelty rests on the claim that latent communication carries information beyond the sampled token sequence. That claim is load-bearing because if false, the method is just an elaborate way to pass a soft summary of the last 64 tokens, which would not explain its gains over full-text transmission. The paper's own evidence for extra information is a single case study (Table 4) showing that a critic can recover the term 'koilonychia' when it is absent from the last K tokens. No quantitative or controlled experiment measures this. The proposed SuffixText control would isolate the contribution of 'extra-message information' from that of the continuous-prefix interface. If StateBridge does not outperform SuffixText, the central contribution must be rephrased: it would show that a soft-prefix presentation of the suffix is more effective than text, not that hidden states preserve more than tokens. This does not change the verdict from CONDITIONAL: the paper is a solid candidate but needs this experiment (and ideally error bars) before the strong form of the claim is accepted.","tokens_in":15962,"tokens_out":9440,"duration_ms":104089,"concrete_test":"Add a 'SuffixText' baseline to Table 1: instead of the aligned prefix, prepend the decoded text of the last K generated tokens to the receiver's prompt, keeping K=64 and all pipeline and decoding settings fixed. If StateBridge's average accuracy minus SuffixText's is not positive and material (e.g., at least 2 points) across the 26 model-task pairs, then there is no evidence that the prefix conveys information beyond the last K tokens, and the central claim would need to be weakened. Run on the same A100 setup with identical seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution (Contribution 1) is that closed-form alignment enables latent communication that preserves information beyond discrete tokens. The mechanism depends on the unproven assumption that the final K hidden states of a generated message, after alignment and anchoring, encode the entire message, including content from tokens earlier than the last K. Section 3.2 asserts this via 'subsequent hidden states aggregate information from the preceding context', and Appendix A.1 only bounds text capacity. The only direct evidence is the qualitative Table 4 case study (one MedQA example, K=16/64/128). No systematic experiment measures how much of the earlier message is recoverable from the prefix, or whether downstream gains require that extra information. The baselines are full-text (TextMAS) and KV-cache (LatentMAS), not a last-K-token text prefix; thus the observed improvements could in principle come from the soft-prefix interface (e.g., better conditioning or format) rather than from transmitting more than the suffix. The paper even notes at K=128 'the aligned prefix no longer carries continuous information beyond the suffix tokens', showing the effect is fragile. If the prefix only summarizes the last K tokens, StateBridge's advantage over full-text communication would disappear, contradicting the headline claim. The reader's weakest assumption captures this, but it is tested only indirectly through downstream accuracy and one case study.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces StateBridge, a training-free protocol for latent communication between homogeneous LLM agents. The sender's last K final-layer hidden states are centered, whitened, and rotated to match the embeddings of the last K decoded tokens via an orthogonal Procrustes solution; the result is norm-calibrated, partially anchored to nearest vocabulary embeddings, and prepended as a continuous prefix to the receiver's prompt. The authors evaluate the method on math, QA, and code benchmarks with Qwen3-4B/8B/32B and OLMo3-7B-Think, reporting best or tied-best on 22 of 26 model-task pairs, with ablations over components and a qualitative case study. The central claim is that closed-form alignment alone resolves the hidden-state/embedding mismatch and permits effective training-free latent communication.","tokens_in":16201,"tokens_out":15808,"duration_ms":170837,"significance":"Training-free, portable latent communication is a valuable goal, and StateBridge's per-message closed-form alignment is a clean idea that avoids trained projectors and layer-wise KV injection. The paper's strengths include a formal derivation of the Procrustes solution, a span-confinement result for ridge regression, a careful ablation isolating the rotation, norm calibration, and vocabulary anchoring, and a random-noise control that rules out a trivial prefix effect. If the empirical claims survive re-evaluation, the method would strengthen the case that continuous inter-agent channels can be built from off-the-shelf models at negligible cost. The significance is therefore substantial, but the current evidence does not yet nail the mechanism: the 'beyond the suffix tokens' claim rests on a single uncontrolled example, and all head-to-head results are single runs.","major_comments":[{"comment":"All evaluations in Table 1 are single runs with stochastic decoding; no standard deviations, seeds, or significance tests are reported. Many decisive comparisons are small (e.g., Qwen3-32B ARC-C is a 1.2-point loss and GSM8K losses are 2.5-2.6 points; Qwen3-4B ARC-C gain is 1.4 points). With run-to-run variance at this scale, the headline '22 of 26' and the conclusion that StateBridge 'consistently outperforms' the strongest baseline are not yet supported. Please report repeated runs with variance estimates or statistical tests for the main comparisons.","section":"Section 5, Table 1"},{"comment":"The paper's load-bearing assumption is that the final K hidden states encode the full message, not just the last K tokens. The only direct evidence is the case study in Table 4, which is a single MedQA example; the critic receives the question as text in addition to the prefix, so part of the 'recovery' could be inference from the question rather than from the prefix. The paper itself notes in Section 6.4 that at K=128 'the aligned prefix no longer carries continuous information beyond the suffix tokens.' A control that transmits only the last-K-token text, or a control that directly uses the suffix embeddings R as the prefix, would separate information recovered from earlier context from a generic soft-prefix benefit. Without such a control, the improvement over TextMAS could be due to the interface rather than to recovering content beyond the suffix.","section":"Section 3.2, Table 4"},{"comment":"Proposition A.3 shows only that the rotation step preserves the Gram matrix of the whitened states; the final injected prefix also undergoes dewhitening, norm calibration, and vocabulary anchoring (Eqs. 7-9), none of which is an isometry. Main-text statements such as 'Orthogonal alignment preserves these semantic similarity' (Section 6.1) therefore overstate the guarantee used to explain the code-generation gains. Please state in the main text that the isometry holds only for the rotation in whitened coordinates, and either adjust the explanatory claims or provide an additional result showing the later steps do not materially alter the relevant geometry.","section":"Section 3.3, Appendix A.3"}],"minor_comments":[{"comment":"Please report the number of independent runs and random seeds used for all main tables; this is needed to assess the stability of the reported accuracies and pass@1 scores.","section":"Appendix B.3"},{"comment":"The 'Random Noise' row should specify how the noise vectors are generated and whether they are norm-calibrated and vocabulary-anchored; if not, the comparison conflates information content with scale and compatibility.","section":"Table 2"},{"comment":"Please state the decoding temperature and sampling protocol for the critic's recovery in the case study, since this is the only direct evidence for the information-preservation mechanism.","section":"Table 4"},{"comment":"The additional benchmarks AIME24, AIME25, and GPQA are reported only for Qwen3-8B and Qwen3-32B; consider reporting OLMo3-7B-Think on these tasks to strengthen the cross-family generalization claim.","section":"Section 4"},{"comment":"Calling the whole interface a 'closed-form orthogonal transformation' is imprecise: the full pipeline includes whitening, norm calibration, and vocabulary anchoring, and only the Procrustes step is an orthogonal map.","section":"Abstract and Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well written and the derivations are sound; the proposed idea is timely and potentially significant. The main gaps are empirical rigor (single runs on close margins) and the absence of a suffix-only or suffix-embedding control that would validate the central information-preservation claim. Both are addressable within the paper's scope, so I would not recommend rejection. Please also ensure the case study is not over-interpreted, since the critic receives the question as additional text input."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful and mostly honest paper. The new thing is that you can take final-layer hidden states from the sender, align them to the receiver's embedding space with a closed-form Procrustes rotation plus norm calibration and vocabulary anchoring, and inject them as a continuous prefix—no training, no KV-cache surgery. It beats text handoff and LatentMAS on 22 of 26 model-task pairs, and the ablations show each component earns its keep. The ridge-regression comparison and the span-confinement lemma in Appendix A.2 are nicely done: they give a real geometric reason why the orthogonal constraint matters. Code is promised. The math is correct as far as I can tell.\n\nWhere it wobbles: the paper's Contribution (1) says closed-form alignment resolves the mismatch and enables effective latent communication that carries information beyond the sampled tokens. That last part is the load-bearing claim. The evidence is one MedQA case study where a critic recovers 'koilonychia' from a K=16 prefix, plus the assertion that later hidden states aggregate earlier context. There is no systematic test of how much of the full message is recoverable, and the baselines are full-text and KV-cache, not a last-K-token text prefix. So the gains could in principle come from the soft-prefix interface itself (better conditioning, format) rather than from transmitting more than the suffix. The paper even admits at K=128 the prefix stops carrying continuous information beyond the suffix; that's honest, but it highlights how fragile the claim is. A simple control—send the receiver only the last K tokens as text, or text plus an aligned prefix of those same tokens—would settle it. The absence of error bars or multiple seeds is a real but standard weakness; the hyperparameters (K, alpha) are chosen on MedQA and then frozen, which is a minor selection effect, not a loophole.\n\nBottom line: if you work on multi-agent LLM systems or continuous prompting, this is worth a read. The method is likely to be useful and the paper is worth engaging with. But the central scientific claim is overstated relative to the evidence. If I were handling it, I'd send it to review and ask for the suffix-text control and at least a few seeds. That's a revise-and-resubmit, not a reject.","headline":"A clean, well-ablated method for training-free latent communication whose load-bearing claim about transmitting more than the suffix is backed by one case study, not a control.","tokens_in":16710,"tokens_out":2633,"would_cite":true,"duration_ms":27239,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training-free latent communication lets LLM agents beat text baselines on 22 of 26 tasks.","keywords":["latent communication","hidden-state alignment","orthogonal Procrustes","multi-agent LLM systems","training-free","continuous prefix","embedding-space compatibility","KV-cache transfer"],"falsifier":"Take a task where the sender's message is longer than K tokens and the last K tokens are deliberately uninformative (e.g., a repeated filler word). If StateBridge's receiver still recovers the earlier content, the prefix carries information beyond the suffix. If it cannot, the claim that the final K states encode the whole message fails. Also, run the same setup with an isotropic random rotation instead of the Procrustes rotation; if performance stays high, the isometry and alignment-specific SVD is not doing the work.","tokens_in":15772,"feed_emoji":"🧠","tokens_out":1528,"duration_ms":17424,"temperature":0.7,"pith_summary":"This paper claims that off-the-shelf LLM agents can communicate through continuous hidden states instead of text, without any training or architectural changes. The key is a closed-form alignment that rotates a sender's final-layer hidden states into the receiver's input embedding space, then calibrates norms and anchors the vectors near real vocabulary embeddings. On math, code, and QA benchmarks across four models, this beats text-based and KV-cache baselines on most tasks, suggesting that the main barrier to latent communication is geometric compatibility, not model training.","feed_headline":"Training-free latent chat lifts LLM agents past text on 22 of 26 tasks","feed_subtitle":"Rotating hidden states into the receiver's embedding space preserves the info that token sampling discards.","key_machinery":"The central object is the closed-form alignment interface: a three-step transformation that maps a sender's final-layer hidden states S to an aligned prefix that the receiver can read. It starts with centering and whitening both S and the reference token embeddings R, then solves an orthogonal Procrustes problem Q* = argmin_Q ||S_w Q - R_w||$_F^{2}$ with Q^T Q = I, whose solution is the SVD product UV^T. This orthogonal rotation preserves pairwise distances and angles in the whitened space. Two compatibility steps follow: norm calibration rescales each vector to the average vocabulary-embedding norm, and vocabulary anchoring moves each vector a fraction $\\alpha$ toward its nearest vocabulary embedding. The paper proves that replacing the orthogonal map with ridge regression confines the prefix to the span of the sampled token embeddings, destroying the continuous information that makes latent communication worthwhile.","core_discovery":"The central claim is that a closed-form orthogonal alignment alone resolves the representation mismatch between final-layer hidden states and input embeddings, enabling training-free and effective latent communication between off-the-shelf LLM agents. The method, StateBridge, extracts the final K=64 hidden states of a sender's generated message, aligns them to the corresponding token embeddings via centering, whitening, and orthogonal Procrustes rotation, then adjusts norms and blends slightly toward vocabulary embeddings. The aligned prefix is prepended to the receiver's prompt embeddings. On 26 model–task pairs spanning Qwen3-4B/8B/32B and OLMo3-7B, StateBridge achieves the best or tied-best score on 22 pairs, improving average accuracy by 2.4–2.9 points over the strongest baseline. Ablations show that the orthogonal constraint is what preserves the geometry that carries information beyond the sampled tokens.","pith_inferences":["One testable extension the paper leaves implicit: if the alignment preserves geometry, then the same closed-form interface could align hidden states between different models of the same family or even across families, since it only needs a shared embedding dimension and a reference matrix R.","The theory suggests that ridge-regression alignment should be strictly worse in any task where the receiver must reconstruct hidden reasoning structure; a direct experiment varying the whitening strength or using anisotropic scaling could further test the isometry hypothesis.","The norm-calibration step implies that attention-dot-product scale is a dominant failure mode for raw hidden-state transfer; a cheap diagnostic would compute attention entropy of the receiver on aligned vs. unaligned prefixes.","One could integrate StateBridge with learned communication by using the orthogonal alignment as a warm-start, potentially reducing projector training data while keeping portability."],"forward_implications":["If the claim holds, any homogeneous multi-agent system can upgrade its inter-agent channel to a training-free continuous prefix, with no model updates and no learned projectors.","Latent communication becomes portable across model families: because StateBridge touches only the input embedding layer, it transfers to architectures where KV-cache injection degrades (e.g., OLMo3-7B).","The geometric-preservation result implies that the information channel between agents is no longer bounded by the discrete token capacity K*log2(V) bits; continuous states can carry confidence and alternative-reasoning traces.","The method's gains concentrate on harder benchmarks (GPQA, AIME24/25, code generation), suggesting that text handoffs are most lossy precisely where nuanced intermediate reasoning matters.","The ablation results indicate that simply injecting raw hidden states or arbitrary continuous vectors is not enough; the prefix must be both informative and compatible, which localizes the design principle for future latent-communication interfaces."],"supporting_citations":[{"why":"Provides the orthogonal Procrustes solution that is the core of the alignment interface.","marker":"[Schönemann, 1966]"},{"why":"Defines the LatentMAS baseline, the four-agent pipeline, and the evaluation protocol that StateBridge follows and outperforms.","marker":"[Zou et al., 2026]"},{"why":"The Cache-to-Cache KV-cache transfer method that StateBridge compares against as a latent baseline.","marker":"[Fu et al., 2026]"},{"why":"An embedding-based latent communication method requiring trained projectors, which StateBridge contrasts as limited in portability.","marker":"[Du et al., 2026]"},{"why":"ThoughtComm, another trained-projector latent method that StateBridge positions against.","marker":"[Zheng et al., 2025]"},{"why":"Supplies the geometric assumption that pairwise proximity in hidden-state space encodes semantic similarity, which motivates the orthogonal constraint.","marker":"[Ethayarajh, 2019]"}],"fun_headline_variants":["Hidden-state chat beats text for LLM agents on 22 of 26 tasks","Train-free hidden-state link lifts LLM agents on 22 of 26 tasks","LLM agents converse via hidden states, no training, 22/26 wins","Orthogonal alignment lets LLM agents share hidden states, beat text 22/26"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The final K hidden states of a generated message, after a single global rotation to the embedding space, encode enough of the entire message that the receiver can act on it. If those K states summarize only the last K tokens, the method's advantage over text disappears.","fun_headline_variants_meta":{"raw":{"variants":["Hidden-state chat beats text for LLM agents on 22 of 26 tasks","Train-free hidden-state link lifts LLM agents on 22 of 26 tasks","LLM agents converse via hidden states, no training, 22/26 wins","Orthogonal alignment lets LLM agents share hidden states, beat text 22/26"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000643,"raw_usage":{"total_tokens":2941,"prompt_tokens":915,"completion_tokens":2026,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":1938}},"tokens_in":531,"tokens_out":2026,"duration_ms":15524,"temperature":1.0,"reasoning_tokens":1938,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:45:44.988337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a task where the sender's message is longer than K tokens and the last K tokens are deliberately uninformative (e.g., a repeated filler word). If StateBridge's receiver still recovers the earlier content, the prefix carries information beyond the suffix. If it cannot, the claim that the final K states encode the whole message fails. Also, run the same setup with an isotropic random rotation instead of the Procrustes rotation; if performance stays high, the isometry and alignment-specific SVD is not doing the work.","supporting_citations":[{"cited_title":"How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings","cited_arxiv_id":null,"evidence_quote":"Supplies the geometric assumption that pairwise proximity in hidden-state space encodes semantic similarity, which motivates the orthogonal constraint."}],"review_version":1}