{"id":"38382913-662d-445d-afbe-bec80534a161","arxiv_id":"2602.03695","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Multi-agent systems built from three KV-cache-communicating primitives plus an organizing agent beat single agents and text-based multi-agent baselines on eight benchmarks at lower cost.","lead":"The paper proposes turning multi-agent AI systems into reusable \"primitives\"—Review, Voting and Selection, and Planning and Execution—that talk through the model's hidden key-value memory instead of natural language. If the approach holds up, it could make multi-agent LLM systems cheaper, faster, and more robust to long or noisy conversations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2)'s KV-cache splicing skips cross-attention between the spliced blocks; the paper calls it an assumption but never validates it, so the latent-communication claim is unsupported.","rationale":"The reader's weakest_assumption identifies exactly the issue I would raise: Eq. (2) is the theoretical foundation for KV-cache communication, and it is asserted without validation. I agree that this is the most load-bearing concern. The rest of the paper—broad empirical gains across eight benchmarks and five backbones, stress tests showing robustness to noise, ablations on RoPE and the Organizer—provides real but indirect evidence. However, none of that evidence validates the specific equivalence in Eq. (2), because the experiments compare KV-cache communication to text-based message passing, not to the oracle of processing the concatenated token sequence. The stated condition for Eq. (2) ('both agents share the same model parameters and positional encoding scheme') is insufficient: shared parameters and RoPE re-indexing fix the positional alignment but not the missing cross-attention from A's tokens to B's system prompt. This is not a question of consensus; it is a correctness risk in the core mechanism. Secondary issues—no released code, no error bars, and the abstract's headline range overstating the largest-backbone results—are real but do not change the central diagnosis. The appropriate disposition remains CONDITIONAL: the paper should be accepted only if the authors either validate Eq. (2) with a direct oracle comparison or weaken the claim to a strictly empirical robustness property. Since the reader already reached CONDITIONAL on the same basis, my recommendation is UNCHANGED.","tokens_in":25240,"tokens_out":5689,"duration_ms":68761,"concrete_test":"On Qwen3-8B, sample 100–200 examples from AIME25/GSM8K. For each primitive, record the actual A-token sequence s_A. Build an oracle context [x_B; s_A] and decode y from the same LLM in a single forward pass. Compare oracle accuracy and output distribution (mean per-token log-likelihood, token agreement rate) with the KV-splice implementation, plus a third condition decoding y from s_A alone. If the KV-splice distribution matches the s_A-alone condition rather than the oracle, Eq. (2) is empirically falsified; if it matches the oracle, the assumption is supported. A complementary analytical check: in a one-layer linear attention model, compute p(y|[x_B;s_A]) exactly and compare with the splice distribution for generic weights to show they differ.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is Eq. (2): p(y|Z_A, x_B) ≈ p(y|s_A, x_B) when parameters and position encodings are shared. This is stated as an assumption and is not derived or empirically checked. It is also not generally valid. In the true sequence [x_B; s_A], the tokens of s_A are processed with self-attention that includes the x_B tokens, so their layer-ℓ key/value states are functions of x_B. In the splice, the stored Z_A was produced by agent A attending only to [x_A; y_A]; the B system KV were produced without later context, but by causality that part is unchanged. Parameter sharing and RoPE re-indexing fix positions only; they cannot inject x_B into the hidden states that produced A's cached K/V. Hence the spliced cache is not the cache of the concatenated token sequence in general. The §3.2 stress tests (Tables 1–2) do not settle this: they compare KV-cache communication with natural-language message passing under injected noise or long context, not with the oracle of decoding from the concatenated text. Observed gains could therefore stem from a different mechanism, e.g., avoiding noisy textual message passing, rather than from the claimed latent equivalence. Since all three primitives and the Organizer composition rely on this splice, Eq. (2) is the load-bearing premise for the reusable-latent-building-blocks claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Agent Primitives, a set of three reusable latent building blocks (Review, Voting and Selection, Planning and Execution) for LLM-based multi-agent systems. Primitives communicate internally via KV-cache splicing rather than natural language, and an Organizer agent composes them per query using a knowledge pool of prior successful structures. The authors claim consistent accuracy improvements over single-agent and text-based MAS baselines (12.0–16.5% average gains over single agents), roughly 3–4× reductions in token usage and latency relative to text-based MAS, and modest 1.3–1.6× overhead over single-agent inference, across eight benchmarks and five backbones.","tokens_in":25575,"tokens_out":5391,"duration_ms":52264,"significance":"If the results hold, the work is a valuable step toward task-agnostic, modular MAS design. The core idea — abstracting MAS into reusable latent primitives — is well-motivated and the evaluation is broad (8 benchmarks, 5 backbones, 3 primitive variants, ablation studies for the Organizer, knowledge pool, and RoPE). The stress-test experiments in §3.2 provide intuitive evidence that latent communication is more robust to noise and long context than natural-language message passing. However, the central theoretical assumption underlying KV-cache communication is not validated, several accuracy/efficiency claims are overstated or contain internal inconsistencies, and the reported numbers lack statistical support. These issues are fixable and do not necessarily invalidate the empirical approach, but they need to be addressed before the central claims can be accepted.","major_comments":[{"comment":"The Input-Output Alignment Assumption is load-bearing but not validated. The spliced KV cache is not the cache that would be produced by decoding from the true concatenated sequence [x_B; s_A], because A's keys/values were computed with self-attention only over A's own tokens; they never attend to x_B in the intermediate layers. RoPE re-indexing fixes positions but cannot inject x_B's context into A's cached states. The stress tests in §3.2 (Tables 1–2) compare KV-cache communication against natural-language communication under injected noise and long context, not against the oracle of decoding from the concatenated text. Consequently, the observed gains could stem from avoiding noisy textual message passing rather than from the claimed latent equivalence. Please provide a direct validation of Eq. (2), e.g., compare decoding from the spliced cache versus decoding from the full concatenat","section":"§3.1, Eq. (2)"},{"comment":"The text states that primitives-based MAS outperforms existing MAS methods across all benchmarks. This is factually incorrect: on GSM8K, Self-Consistency achieves 95.0% while Primitives-based MAS achieves 93.8%. This discrepancy undermines the 'across all benchmarks' claim. Please revise the claim to acknowledge the exception, or provide a corrected table. The GSM8K result also raises a question about whether the advantage is consistent or task-dependent.","section":"Table 4, §5.3"},{"comment":"The abstract's claim of '12.0–16.5% average accuracy improvement' is not representative of all five evaluated backbones. Table 3 reports average improvements of +6.6% for DeepSeek-R1-Distill-Qwen-32B and +6.3% for DeepSeek-R1-Distill-Llama-70B. If the range is meant to apply only to a subset of models (e.g., Qwen3-8B and Qwen3-14B), this must be stated explicitly. Otherwise, report the full observed range with appropriate caveats.","section":"Abstract and §5.2"},{"comment":"All accuracy numbers appear to be single-run point estimates, with no error bars, confidence intervals, or significance tests. Many comparisons involve small differences (e.g., Table 3, GSM8K: +0.0 to +1.4pp on DeepSeek-R1-Distill-Qwen-32B; Table 4, several baselines). Without repeated runs or statistical testing, the rank ordering and the 'consistent improvement' claim are not supported. Please report standard deviations across multiple runs (or seeds) and perform appropriate significance tests where differences are small.","section":"Tables 3, 4, 8"},{"comment":"The TextMAS row for AIME24 reports 3,455 tokens with a −60.5% change relative to single-agent, which contradicts the claim that text-based MAS is substantially more token-hungry. This appears to be a data error or typo, and it affects the credibility of the token-efficiency analysis (and the abstract's 3–4× reduction claim). Please verify all entries in Tables 9 and 10; Table 10 also contains a duplicated model label ('DeepSeek-R1-Distill-Qwen-32B' appears twice, likely the second block should be 'DeepSeek-R1-Distill-Llama-70B').","section":"Table 9, DeepSeek-R1-Distill-Qwen-32B row"},{"comment":"The Knowledge Pool is seeded with structures from the same MAS frameworks that are then compared in Table 4 (e.g., Multi-Agent Debate, DyLAN, Self-Refine, AFlow, MAS-GPT). This gives the Organizer prior access to configurations of the baselines, which may bias the comparison in favor of primitives-based MAS. The paper should explicitly acknowledge this overlap and discuss its implications, ideally by also comparing against baselines that are not represented in the Knowledge Pool.","section":"§4.3, Knowledge Pool and Table 4"}],"minor_comments":[{"comment":"The primitive name is consistently misspelled as 'V oting' (e.g., Figure 1, Tables 3, 7, 8). Change to 'Voting'.","section":"Throughout"},{"comment":"The reference to 'MacNet (Qian et al.)' lacks a year and venue. Please complete the reference.","section":"§2, Related Work"},{"comment":"The stopping condition is described as 'derived from the intermediate latent states' but no concrete mechanism is given. A brief clarification would help reproducibility.","section":"§4.2, Review Primitive"},{"comment":"Using GPT-5.2 as the default Organizer introduces a proprietary, potentially expensive component. The paper should note the cost/access implications and whether the Organizer could be replaced by an open-weight model without major degradation (the Claude-4 ablation is a good start).","section":"§5.1, Implementation Details"},{"comment":"The caption says 'Qwen3-4B and 8B' but the table contains 4B, 8B, and 14B rows. Please correct the caption.","section":"§C, Table 8"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea and a broad evaluation, but the central KV-cache equivalence assumption is unverified and several claims are overstated or contain data errors. The authors should be encouraged to address these issues; the empirical direction is worth pursuing. In particular, the Table 4 GSM8K inconsistency and the Table 9 token anomaly should be corrected before any further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — this one is worth a serious read, but keep the central claim labeled. The genuinely new bit is treating recurring MAS interaction patterns (review, vote/select, plan/execute) as plug-in primitives that talk through KV caches, plus an organizer that picks and composes primitives from a small knowledge pool. That is a real conceptual step beyond one-off role/prompt engineering, and it gives other people a vocabulary. The evaluation is broad: eight benchmarks, five backbones, existing MAS comparisons, ablations. I believe the relative ordering over text-based MAS, and the efficiency numbers (3×–4× cheaper than text MAS) are plausible and useful.\n\nThe soft spot, as you'd expect, is Eq. (2). The paper says conditioning on A's KV cache is approximately equivalent to conditioning on A's full token sequence when parameters and positional encodings are shared. It is not derived and not validated. The stress-test critique is correct: a cache produced by A in isolation never had B's system tokens in its self-attention windows, so splicing it after B's system prompt is not the cache of the concatenated sequence. Re-indexing RoPE fixes positions, not cross-attention. All three primitives and the organizer rely on this splice, so this is load-bearing, and the §3.2 stress tests don't certify it: they compare KV communication to noisy text message passing, not to decoding from the concatenated text. Observed gains may partly come from avoiding noisy text, which is valuable but not the same as latent equivalence.\n\nAlso: no code or seeds, no error bars or significance tests on the main tables, and the abstract's 12.0–16.5% gain range is an average over small and large backbones; on the two largest models, gains are 6.3–6.6%, so the headline slightly overstates. The knowledge pool is seeded from the same MAS families it later beats in Table 4; that's mild contamination, not fatal, because the pool only guides structure. I'd name it and move on.\n\nWho gets value from this: anyone working on reusable abstractions for LLM agents or on latent communication in multi-agent systems. It deserves a serious referee. The authors need to either validate or weaken Eq. (2), ideally with a decoding-from-concatenated-text oracle, report variance, and release code. If they fix those, this could be a solid paper.","headline":"Useful abstraction, broad evaluation, but the KV-cache equivalence that carries the whole latent-communication story is asserted, not validated; referee it, but expect major revision.","tokens_in":26065,"tokens_out":1801,"would_cite":true,"duration_ms":20789,"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":"Three reusable latent primitives replace hand-built agent teams, communicating by key-value cache.","keywords":["multi-agent systems","KV cache","latent communication","LLM agents","reusable primitives","agent composition","review voting planning","efficiency"],"falsifier":"Run the same task twice: once with agent B given the full text of agent A's output followed by its own system prompt, and once with agent B given the spliced KV cache of both. Choose a task where the correct answer depends on details early in A's output (for example, an instruction injected at the beginning of a long context). If accuracy or instruction-following differs systematically between the two conditions, the alignment assumption fails and the latent channel is lossy.","tokens_in":25131,"feed_emoji":"🧩","tokens_out":5623,"duration_ms":53899,"temperature":0.7,"pith_summary":"This paper tries to establish that hand-designed multi-agent LLM systems can be replaced by a small catalog of reusable computation patterns—review, voting-and-selection, and planning-and-execution—that exchange information through key-value (KV) caches instead of natural language. The claim is that many existing multi-agent architectures reduce to these patterns, and that packaging them as primitives with a standard agent interface lets an Organizer compose them per query with no manual role or prompt engineering. If correct, multi-agent system design becomes modular and task-agnostic, and inter-agent communication stops degrading in long-context, noisy settings. Across eight benchmarks and five model families, the paper reports 12.0–16.5% higher average accuracy than a single agent, roughly 3–4x lower token use and latency than text-based multi-agent systems, and only 1.3–1.6x overhead over single-agent inference.","feed_headline":"Three reusable agent primitives beat hand-built multi-agent teams","feed_subtitle":"Latent key-value communication lifts accuracy 12-16% and cuts tokens and latency ~3-4x vs text-based agents.","key_machinery":"The load-bearing mechanism is KV-cache splicing: the final key-value cache produced by agent A is concatenated, layer by layer, with the system-prompt cache of agent B, and agent B then autoregressively decodes from that combined cache. This replaces textual message passing with a single continuous latent context. The validity of the mechanism rests on the Input-Output Alignment Assumption, p(y | Z_A, x_B) ≈ p(y | s_A, x_B), that conditioning on cached keys/values is distributionally equivalent to conditioning on the full token sequence. Because modern transformers use rotary positional encodings, the spliced positions from agent A must be re-indexed by the length of agent B's system prompt.","core_discovery":"On its own terms, the paper's central discovery is that latent, cache-based communication is not just a cheaper substitute for text between agents; it is the mechanism that makes reusable multi-agent computation possible. A transformer agent that finishes a computation leaves behind a key-value cache—the per-layer attention keys and values it stored during decoding—and the paper treats this cache as a communicable latent state. Concatenating one agent's cache with another agent's system prompt lets the second agent continue decoding without ever seeing the first agent's text. The three primitives (Review, Voting and Selection, Planning and Execution) are the recurring internal patterns that","pith_inferences":["If the alignment assumption holds beyond the tested settings, KV-cache splicing becomes a general-purpose inter-agent transport, so new primitives (memory, hierarchical decomposition, tool use) could be added without redesigning communication.","A direct paired test—same prompt, full-text conditioning versus spliced-cache conditioning—would isolate whether observed gains come from the latent channel itself or from the particular primitives and prompts; the paper does not report this control.","The stress-test results suggest that text serialization, not model capability, may be the main bottleneck in long-horizon multi-agent reasoning; if so, latency budgets in deployed agent systems could be repurposed toward more primitive passes.","The 45-structure knowledge pool could be grown online as queries are solved, which would let the Organizer improve its composition choices over time and eventually eliminate the need for seed structures."],"forward_implications":["A multi-agent system can be assembled automatically from a small fixed set of primitives, with an Organizer choosing the composition per query, rather than by hand-coding roles and prompts.","KV-cache communication is more resistant than text to long-context dilution and injected noise; in the paper's stress tests, compliance and accuracy degrade far less when messages are latent.","Primitives-based MAS improve average accuracy by 12.0–16.5% over single-agent baselines on math, code, and Q&A, with gains that persist across five model backbones.","Compared with text-based MAS, primitives cut token usage and inference latency by roughly 3–4x, so the accuracy gains come with lower communication cost.","Composing several primitives adds another 3.5–7.0% over the strongest single primitive, indicating that primitive composition itself contributes beyond individual patterns."],"fun_headline_variants":["Latent primitives: 12-16% accuracy lift, 3-4x less tokens","Reusable agent blocks via key-value cache cut cost 3-4x","Three primitives deliver 12-16% accuracy gain over single agents","Organizer composes latent primitives for more robust multi-agent AI","Key-value cache communication makes multi-agent systems 3-4x cheaper"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole system stands or falls on the assumption that a key-value cache is a lossless stand-in for the token sequence that created it; if the cache loses or distorts information when spliced next to another agent's system prompt, every latent message between agents is corrupted.","fun_headline_variants_meta":{"raw":{"variants":["Latent primitives: 12-16% accuracy lift, 3-4x less tokens","Reusable agent blocks via key-value cache cut cost 3-4x","Three primitives deliver 12-16% accuracy gain over single agents","Organizer composes latent primitives for more robust multi-agent AI","Key-value cache communication makes multi-agent systems 3-4x cheaper"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001059,"raw_usage":{"total_tokens":4311,"prompt_tokens":808,"completion_tokens":3503,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":3411}},"tokens_in":552,"tokens_out":3503,"duration_ms":25776,"temperature":1.0,"reasoning_tokens":3411,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:51:15.909313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same task twice: once with agent B given the full text of agent A's output followed by its own system prompt, and once with agent B given the spliced KV cache of both. Choose a task where the correct answer depends on details early in A's output (for example, an instruction injected at the beginning of a long context). If accuracy or instruction-following differs systematically between the two conditions, the alignment assumption fails and the latent channel is lossy.","supporting_citations":[],"review_version":1}