{"id":"0be85876-382c-402b-9e77-4ee57774f4a7","arxiv_id":"2506.02739","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper argues that natural language is structurally mismatched to LLM internal representations, so future AI agents should abandon it for inter-agent communication and train models with structured communication primitives.","lead":"Natural language is a lossy and fragile medium for communication between AI agents, because it cannot fully carry the high-dimensional internal states of large language models. The paper argues for a new model-building paradigm in which agents exchange structured, tensor-based messages and are trained explicitly for multi-agent coordination.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed non-zero lower bound on semantic loss is not derived; finite-rate channel lossiness alone does not imply task-relevant drift because a learned inverse can recover relevant state.","rationale":"The reader identified as the weakest assumption the asserted non-invertibility of f and the consequent inevitability of cascading loss. My stress-test sharpens this: the paper does not prove that the finite-rate channel loss is task-relevant, nor that a learned approximate inverse cannot bring task-relevant error to zero. This is the same load-bearing assumption, and the reader's conditional verdict already captures the resulting correctness risk. The paper is best read as a research agenda rather than a demonstrated theorem; the formal apparatus in Section III.A is a formalization of the thesis, not evidence for it. I agree with the reader's verdict and find no further concern that would move it.","tokens_in":14866,"tokens_out":5236,"duration_ms":54921,"concrete_test":"Construct a synthetic two-agent referential coordination task with a K-dimensional continuous state; fix a small token budget; train or fine-tune two LLM agents to communicate via free-form English messages and measure both task success and state-reconstruction error. If fine-tuned agents achieve near-perfect task success for K exceeding the naive upper bound implied by token-budget times log|V|, the inevitable-loss claim fails. More directly, independently re-derive Eq. (3) from an explicit generative process where the receiver is an LLM with a learned inverse g, and compute E[||P(h - g(f(h)))||^2] for the task-relevant projection P; if this quantity tends to zero as message length grows, the non-zero lower bound is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central conclusion that natural-language inter-agent communication inevitably incurs cascading semantic loss rests on the assertion that the generation map f: H -> L in Eq. (1) is many-to-one and non-invertible, and that this forces a non-zero lower bound on L_cascade in Eq. (3). This does not follow. Any finite-rate channel from a continuous space is many-to-one; what matters is whether the lost information is task-relevant and whether the receiver's approximate inverse g: L -> H can recover the task-relevant subspace. The paper never proves that the equivalence classes of f cross task-relevant distinctions, nor that g cannot be learned (through shared training or fine-tuning) to keep task error small. Eq. (3) is not a derivation; it defines total error and then asserts that errors accumulate without correction. No argument rules out error-canceling dynamics, targeted re-querying, or messages long enough to quantize relevant state. Thus the claimed non-zero lower bound, and with it the necessity of an architectural change, is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that natural language is structurally misaligned with the high-dimensional vector spaces in which LLMs operate, so using natural language for inter-agent communication causes irreversible information loss, semantic drift, and cumulative coordination errors. The authors formalize the generation process as a projection f: H -> L from an internal semantic space to a discrete token space, claim that f is many-to-one and non-invertible, and use this to assert that the cascading semantic loss L_cascade has a non-zero lower bound. They further argue that the next-token prediction training objective lacks structural support for role persistence, task decomposition, and inter-agent state synchronization, and they propose a native multi-agent modeling paradigm with role persistence, structured tensor communication, state synchronization, and functional decoupling. The paper is primarily an argumentative survey: it contains no new experiments, no machine-checked proofs, and its central formal claims rest on unproven assumptions.","tokens_in":15135,"tokens_out":5746,"duration_ms":58361,"significance":"The question the paper raises is timely and practically important: if natural-language communication between LLM-based agents is fundamentally lossy, then scalable multi-agent systems may require new communication protocols or even new training paradigms. The paper offers a useful synthesis of known failure modes, a clear set of design requirements in Section IV, and a helpful discussion of open challenges in Section IV.C. At the same time, the paper's central formal claim is not established: the alleged inevitability of cascading semantic loss is built into the definition of f rather than derived, and the architectural incompatibility claim is asserted rather than demonstrated. The paper is therefore best viewed as a research agenda or position statement rather than a completed technical result, and its current form overstates the conclusiveness of its own analysis.","major_comments":[{"comment":"The claimed non-zero lower bound on L_cascade is not derived; it is assumed. Eq. (1) defines f as many-to-one, and Eq. (3) defines L_cascade as the sum of per-step state-recovery errors. The text then asserts that because f is many-to-one, the error delta_t accumulates and L_cascade has a non-zero lower bound. This does not follow: any finite-rate channel from a continuous space is many-to-one, and the relevant question is whether the lost information is task-relevant and whether an approximate inverse g: L -> H can recover the task-relevant subspace. No argument is given that the equivalence classes of f cross task-relevant distinctions, nor that g cannot be learned through shared training or fine-tuning. Without such an argument, the conclusion that cascading semantic loss is inevitable is unsupported.","section":"Section III.A, Eqs. (1)-(3)"},{"comment":"The formalization conflates sender and receiver states and assumes the receiver's update is T(f^{-1}(f(h_t))). The error term delta_t is defined as the difference between h_t and its reconstruction from the message, but the claim that this error 'accumulates over time' is an assertion, not a consequence of the equation. If the receiver can maintain a conversation-level state, or if the task depends only on a low-dimensional projection of h, the reconstruction error may be irrelevant or non-accumulating. The paper needs either an explicit counterexample with concrete dynamics or a formal condition under which accumulation is guaranteed; otherwise Eq. (3) merely restates the assumption built into Eq. (1).","section":"Section III.A, Eq. (2)"},{"comment":"The claim that the next-token prediction objective 'lacks structural priors' and therefore cannot support role persistence, task boundaries, or multi-agent dependencies is a non-sequitur. A model trained by next-token prediction can in principle learn to maintain role-consistent states from conversational data; the absence of an explicit inductive bias does not imply that the relevant structure cannot be learned. The paper provides no empirical or theoretical evidence that such structures are unlearnable under the standard objective. This is load-bearing because the architectural incompatibility argument motivates the proposed shift to a new training paradigm; as written, it remains an unsupported assertion.","section":"Section III.C, Eq. (6)"},{"comment":"The 'Protocol-Induced Agent Misbehavior' and 'Architectural Incompatibility' arguments repeatedly rely on the same unproved assertion that f is lossy and non-invertible. The empirical examples (AutoGPT, AgentVerse, Cemri et al.) are suggestive, but they are not controlled comparisons between natural language and an alternative protocol; they are consistent with the paper's thesis but do not establish that the cause is structural misalignment rather than limited context windows, insufficient prompting, or other engineering factors. To make this causal claim load-bearing, the paper would need a controlled study or at least a formal model with clearly stated assumptions and testable predictions.","section":"Section III.B and III.C, Eqs. (4)-(5)"}],"minor_comments":[{"comment":"The paper refers to 'MCP' as 'Multi-Agent Communication Protocol' in the introduction, but Ref. [9] is the Model Context Protocol; the acronym is used inconsistently and should be corrected.","section":"Section II.B and Ref. [9]"},{"comment":"The text attributes a claim about lack of explicit state modeling to 'Microsoft Research [14]', but Ref. [14] is by Laban et al.; the attribution is inaccurate.","section":"Section II.C"},{"comment":"The subscripts in Eq. (2) are inconsistent: the left side is h(i)_{t+1} while the right side uses h(j)_t, and the roles of i and j are not defined. Please clarify the state-update notation.","section":"Section III.A, Eq. (2)"},{"comment":"The statement that L_cascade 'admits a non-zero lower bound and cannot be eliminated under standard decoding assumptions' is not accompanied by a precise statement of what those 'standard decoding assumptions' are; the assumptions should be stated explicitly or the 'theoretically' claim should be removed.","section":"Section III.A"},{"comment":"The direct-sum decomposition S = \\bigoplus S^{(i)} with T(S^{(i)}) subset of S^{(i)} is introduced as an 'ideal' condition, but no justification is given for why such a decomposition is necessary or achievable, and the direct-sum notation is never defined.","section":"Section III.C"},{"comment":"Several claims are supported by citations to blog posts and GitHub issues (e.g., Refs. [16], [55], [56]); these are less authoritative than peer-reviewed sources and should be supplemented or clearly flagged as anecdotal.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper reads as an extended position statement rather than a technical result. The formal section is rhetorically strong but mathematically weak: the central 'theorem' restates its own assumption, and the architectural claims are asserted rather than proven. I would not reject on novelty grounds, since the question is timely and the proposed design requirements are useful, but the paper should either substantially weaken its formal claims or replace them with a rigorous argument and controlled experiments. The editor may also wish to consider whether the heavy reliance on arXiv preprints, blog posts, and GitHub issues meets the journal's citation standards."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is a paper with a useful research agenda and an unsupported formal core. Read it as a thought piece, not as a theoretical result.\n\nWhat's actually new: the paper puts two known observations together--natural language is a lossy, ambiguous medium for inter-agent communication, and LLMs are trained on next-token prediction without any explicit multi-agent coordination objective--and draws an architectural conclusion: if you want scalable multi-agent systems, you should consider building models whose internal representations natively support role identity, state synchronization, and structured message passing. That's a real and timely claim, and the paper's related-work synthesis is competent. It honestly engages with the empirical failure modes documented by Cemri et al., Han et al., Laban et al., and others, and it proposes concrete design desiderata (role persistence, structured communication, state sync, functional decoupling) that are reasonable starting points.\n\nThe soft spots are in the formal analysis. Section III.A presents equations as if they establish that natural-language communication must produce cumulative, uncorrectable semantic loss, but the argument is circular. The non-zero lower bound on L_cascade is built out of the assumption that f is many-to-one and non-invertible. Many-to-one is true of any finite-rate channel; the real question is whether the lost information is task-relevant and whether a receiver can learn an approximate inverse that recovers the relevant subspace. The paper never proves that f's equivalence classes cut across task-relevant distinctions, and it never rules out error-correcting dynamics, targeted re-querying, or messages long enough to quantize the state. So the claim that L_cascade 'cannot theoretically converge to zero without structural priors' is asserted, not derived. The same overreach appears in the claim that chain-of-thought, function-calling, and memory buffers 'cannot circumvent' the problem.\n\nThe paper also overstates novelty--the claim to be the first systematic analysis is not supported by its own reference list, which contains several prior critiques--and it leans on low-quality citations (blogs, Wikipedia, informal web sources) where peer-reviewed or preprint citable work exists.\n\nNone of this kills the paper's value. The architectural question it raises is important, and the proposal section is a reasonable agenda. But the theoretical wrapping should be either downgraded to 'illustrative' or replaced by a real analysis of when natural-language communication is task-lossy.\n\nI'd send this to review, not desk-reject, with a referee request to make the claims match the evidence. The author should either prove a lower bound under explicit assumptions or present the paper as a position piece with the equations removed. For me, it's a 'maybe' for the reading group--the discussion value is high, but the theory needs to be taken with a grain of salt.","headline":"A useful research agenda that overstates its formal case: the paper's central claim of inevitable cascading semantic loss is asserted, not proven.","tokens_in":15630,"tokens_out":3399,"would_cite":false,"duration_ms":35989,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Natural language is the wrong medium for AI-to-AI coordination, this paper argues.","keywords":["multi-agent systems","large language models","natural language communication","semantic misalignment","cascading semantic loss","role persistence","structured communication","next-token prediction"],"falsifier":"Sample many pairs of distinct internal states that map to the same utterance, feed the utterance to a second model, and check whether the recovered state is closer to the true sender state than chance; if the approximate inverse $\\tilde{f}^{-1}$ recovers the sender state accurately, the many-to-one and non-invertibility premise fails. A second check: measure whether task success in a long-horizon two-agent protocol drops monotonically with conversation length when context windows and memory are held fixed; if no such drop appears, the claimed nonzero lower bound on $L_{\\text{cascade}}$ is not supported.","tokens_in":14690,"feed_emoji":"🤖","tokens_out":9208,"duration_ms":83945,"temperature":0.7,"pith_summary":"AI agents that coordinate by sending natural-language messages are, on this paper's view, pushing their high-dimensional internal states through a narrow, lossy pipe. The authors argue that the semantic space of natural language is structurally misaligned with the vector spaces in which LLMs actually reason, so every exchange compresses away information that cannot be recovered, and the errors accumulate across turns into behavioral drift, role confusion, and broken task chains. Because current LLMs are trained only to predict the next token over linear text streams, the paper concludes that they cannot support scalable multi-agent coordination without an architectural change. The proposed remedy is a native multi-agent modeling paradigm in which communication is structured tensor-state exchange, role identities are bound to internal states, and perception, planning, and coordination are decoupled.","feed_headline":"Natural language fails as an agent-to-agent channel, paper argues","feed_subtitle":"Token-based messages compress LLM internal states, causing semantic drift and coordination collapse across agents.","key_machinery":"The load-bearing object is the generation map $f: \\mathcal{H} \\to \\mathcal{L}$, where $\\mathcal{H}$ is the high-dimensional continuous semantic state space of an LLM and $\\mathcal{L}$ is the discrete space of natural-language token sequences. The paper treats $f$ as a many-to-one, non-invertible compression, so that message transmission is modeled as $h^{(j)}_t \\mapsto f(h^{(j)}_t) \\mapsto \\tilde{f}^{-1}(f(h^{(j)}_t)) \\neq h^{(j)}_t$, and it formalizes the accumulated error as $L_{\\text{cascade}} = \\sum_{t=1}^{T} \\mathbb{E}_{j \\in A}[\\|h^{(j)}_t - \\hat{h}^{(j)}_t\\|^2]$. Everything else—the need for role-space binding $h^{(i)}_t \\sim p(h \\mid r^{(i)})$, structured message tensors $m^{(t)}_{ji} = G(h^{(j)}_t)$, and the coordination graph $\\mathcal{G}_t = (\\mathcal{A}, \\mathcal{R}, \\mathcal{E}_t)$—is presented as the remedy once this loss is acknowledged to have a nonzero lower bound.","core_discovery":"The central claim is that natural language is not merely an inefficient protocol for agent-to-agent communication but structurally the wrong medium. The paper models each agent's internal semantic state as a point in a high-dimensional tensor space $\\mathcal{H}$ and the language it emits as a projection $f: \\mathcal{H} \\to \\mathcal{L}$ into a discrete token space. Because $f$ is claimed to be many-to-one and non-invertible, distinct internal states collapse to identical utterances, and a receiving agent's reconstruction $\\hat{h} = \\tilde{f}^{-1}(l)$ almost never equals the sender's state $h$; this per-round error accumulates as a cascading semantic loss with a nonzero lower bound. The same structural mismatch, the paper argues, explains why LLM agents show goal drift, pseudo-execution, role confusion, and broken task chains, and it cannot be fixed by prompting, memory, or better protocols. The positive thesis is that these failures are architectural and must be addressed by training a new class of native multi-agent models with role persistence, structured tensor communication, inter-agent state synchronization, and decoupled perception, planning, and coordination modules.","pith_inferences":["The paper's compression logic implies the same loss should be measurable in a single agent: a model that states a plan in words and later executes it should show measurable divergence between the original plan state and the state recovered from its own paraphrase.","An empirical test the paper does not run would compare task-completion rates over matched interaction budgets between language-only agent teams and teams sharing latent-state communication channels; a large gap would support the proposal, while no gap would undermine it.","The argument sets a correctness ceiling below communication bandwidth: even arbitrarily large models would keep losing state through the discrete-token bottleneck unless training objectives change, not just parameters.","The proposed paradigm suggests new benchmarks built around multi-turn coordination tasks that vary conversation length and role count and measure state-recovery error rather than only final task success."],"forward_implications":["Long-horizon, multi-role agent teams built on current LLMs will accumulate state divergence, so extended task chains should be expected to degrade even when each individual reply looks coherent.","Prompt tuning, memory buffers, and protocol wrappers can reduce symptoms but cannot remove the underlying loss, because the bottleneck is the generation architecture itself.","A workable fix requires models whose role identifiers are bound to internal states, whose communication runs on structured tensors rather than tokens, and whose action policies are decoupled from language output.","Hybrid interfaces will be needed so that natural language remains a human-debuggable channel while the coordination path operates on state-aligned structured messages."],"supporting_citations":[{"why":"Documents the lost-in-conversation effect in long multi-turn planning, evidence for cumulative context loss.","marker":"[14]"},{"why":"Documents semantic drift and role confusion in multi-role LLM systems, evidence for coordination failure.","marker":"[15]"},{"why":"Shows LLM agents describe completion without executing actions, supporting the action-state decoupling claim.","marker":"[18]"},{"why":"Grounds the idea that communication protocols can emerge from agent populations, a contrast to natural language as default.","marker":"[25]"},{"why":"Shows agents can learn minimal task-coupled communication in a differentiable way, the alternative design reference.","marker":"[41]"},{"why":"Shows similar LLMs develop idiosyncratic semantic boundaries, supporting the claim that shared intent cannot be assumed from language.","marker":"[50]"},{"why":"Provides evidence of reasoning loops and planning failures in natural-language task decomposition.","marker":"[51]"},{"why":"Supplies the symbol grounding problem as background for why discrete symbols fail to capture continuous internal meaning.","marker":"[52]"},{"why":"Argues neural networks and linguistic generalization diverge, supporting the structural mismatch claim.","marker":"[53]"}],"fun_headline_variants":["Natural language fails as AI agent communication","AI agents should ditch human language for coordination","Human language limits multi-agent AI coordination","LLMs need a native tongue for agent coordination","Designer language for AI agents, not natural"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the premise that an agent's internal state cannot be recovered from the language it emits; if models can learn near-invertible mappings from text back to internal state, the claimed cascade of semantic loss would not be inevitable.","fun_headline_variants_meta":{"raw":{"variants":["Natural language fails as AI agent communication","AI agents should ditch human language for coordination","Human language limits multi-agent AI coordination","LLMs need a native tongue for agent coordination","Designer language for AI agents, not natural"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1722,"prompt_tokens":1020,"completion_tokens":702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":636}},"tokens_in":636,"tokens_out":702,"duration_ms":6684,"temperature":1.0,"reasoning_tokens":636,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:16:41.851749+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample many pairs of distinct internal states that map to the same utterance, feed the utterance to a second model, and check whether the recovered state is closer to the true sender state than chance; if the approximate inverse $\\tilde{f}^{-1}$ recovers the sender state accurately, the many-to-one and non-invertibility premise fails. A second check: measure whether task success in a long-horizon two-agent protocol drops monotonically with conversation length when context windows and memory are held fixed; if no such drop appears, the claimed nonzero lower bound on $L_{\\text{cascade}}$ is not supported.","supporting_citations":[{"cited_title":"Emergence of grounded compositional language in multi-agent populations,","cited_arxiv_id":null,"evidence_quote":"Grounds the idea that communication protocols can emerge from agent populations, a contrast to natural language as default."},{"cited_title":"The symbol grounding problem,","cited_arxiv_id":null,"evidence_quote":"Supplies the symbol grounding problem as background for why discrete symbols fail to capture continuous internal meaning."},{"cited_title":"Linguistic generalization and compositionality in modern artificial neural networks,","cited_arxiv_id":null,"evidence_quote":"Argues neural networks and linguistic generalization diverge, supporting the structural mismatch claim."}],"review_version":1}