{"id":"c9bcb9de-1d73-4f7b-a21e-8dceff780804","arxiv_id":"2509.05651","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Orchestrator, an active-inference-inspired feedback system for LLM multi-agent teams, substantially raises maze-solving success rates on medium-difficulty mazes but not consistently on hard mazes.","lead":"A new framework called Orchestrator adds a monitoring and feedback layer to teams of language-model agents, and tests it on maze puzzles. It reports large success-rate gains over solo agents on easy and medium mazes, though the gains are less clear on the hardest mazes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. (2)-(3) define U as negative Shannon entropy, so U <= 0 at every step while the Section 3.2 policy thresholds require U > 0.6; the active-inference loop is inert as written.","rationale":"The reader identified threshold fitting and the validity of the VFE proxy as the weak point. My check sharpens this into an internal inconsistency: in Eqs. (2)-(3), U is the negative of a Shannon entropy, hence never positive, while the Section 3.2 policy branches require U greater than 0.6. This is independent of whether thresholds were grid-searched on the testbed and independent of baseline selection. If confirmed by per-step logging, the central claim that Orchestrator optimizes behavior via active inference is unsupported; the observed improvements could stem from the hand-authored tool interface and prompts. The result could be repaired by correcting the sign of the epistemic term, re-running the evaluation, and releasing code that shows actual U values, but as submitted the paper's formal mechanism is vacuous. I therefore move the reader's CONDITIONAL verdict to REJECT.","tokens_in":17079,"tokens_out":7780,"duration_ms":72993,"concrete_test":"Run the medium 18x18 configuration for the 25 reported runs while logging raw U and C per agent-step using Eqs. (2)-(5) exactly as written; if max(U) is at most 0 in every run, no agent is ever in a High Epistemic Drive category, and the orchestration signal is not driven by active inference. If code is not available, re-implement the formulas from the paper and check whether any threshold pair in Appendix A.5 can be reached by any logged value.","verdict_should_be":"REJECT","load_bearing_attack":"Equation (2) defines Uepistemic(n,t,k) = -H[S_{n,t,k} | S_{n-1,t-1,k-1}], and Eq. (3) computes H as the Shannon entropy of token probabilities, -sum_j p_j(k) log p_j(k). Shannon entropy is nonnegative, so U is non-positive at every step. Yet Section 3.2 assigns the two High Epistemic Drive categories only when U > 0.6, and the low-drive categories when U < 0.6; no agent can ever enter the high-drive branches. Because the accuracy cost C in Eq. (4) is also nonnegative, F = U - C in Eq. (5) is non-positive everywhere. The grid-searched thresholds in Appendix A.5 are therefore not merely overfit; on the stated scale they are unreachable. The dynamic weight modulation in Eq. (7) can respond only to the accuracy proxy, not to epistemic information gain, so the paper's central claim that Orchestrator uses active-inference benchmarks to optimize behavior is not supported by the equations as written. The reported performance differences could in principle come entirely from the prompt and tool scaffolding in Appendices A.7-A.8.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Orchestrator, a multi-agent framework for LLM-based maze solving that combines an orchestrator node, reflective benchmarking, and an active-inference-inspired variational free energy (VFE) score. The VFE score is defined as epistemic uncertainty minus an accuracy cost, and threshold-based performance categories trigger dynamic weight adjustments. Experiments on procedurally generated mazes compare solo agents, FE-benchmark-only agents, and FE-plus-orchestration agents using GPT-4.1-nano and GPT-5-nano, reporting success rates, cost, and step counts. The paper claims that active-inference-driven orchestration substantially improves reliability and scalability, with up to 100% success on medium 18x18 mazes and up to 76.67% on hard 25x25 mazes.","tokens_in":17446,"tokens_out":4692,"duration_ms":42151,"significance":"If the mechanism worked as described, Orchestrator would offer a practical coordination method for lightweight LLMs in long-horizon tasks. The paper includes detailed appendices with prompts, algorithms, maze-generation code, and confidence-interval convergence analysis, and it candidly reports settings where orchestration hurts performance. However, because the core VFE equations as written cannot activate the high-epistemic-drive branches, the active-inference loop is inert; the reported gains are therefore not evidence for the paper's central claim. The threshold grid-search on the same difficulty levels further undermines the independent-evaluation claim. The framework may still work as a prompted multi-agent heuristic, but the paper does not establish that active inference is the operative mechanism.","major_comments":[{"comment":"Uepistemic is defined as the negative Shannon entropy of token probabilities, so it is non-positive at every step, while the policy thresholds require U > 0.6 to activate the High Epistemic Drive categories. Consequently no agent can ever enter those categories, F = U - C in Eq. (5) is non-positive, and the dynamic weight modulation in Eq. (7) can respond only to the accuracy proxy. The active-inference feedback loop is inert as written, so the reported improvements cannot be attributed to the VFE mechanism without a corrected formulation and rerun experiments.","section":"Section 3.2, Eqs. (2)-(3)"},{"comment":"The thresholds ϑ1=0.6 and ϑ2=0.4 were selected by grid search on medium- and hard-difficulty mazes, the same difficulties used in Table 1. This makes the reported success rates on medium and hard mazes partly a result of fitting rather than an independent evaluation. Moreover, Appendix A.5 reports different optimal thresholds for hard mazes (ϑ1=0.9, ϑ2=0.01), so the adopted setting is not even the best for the hard condition.","section":"Section 3.2 and Appendix A.5"},{"comment":"The orchestration component does not consistently improve over FE-benchmarking alone. For GPT-4.1-nano on hard mazes, success drops from 84.62% (95% CI 66.47-93.85) to 71.88% (95% CI 54.63-84.44); for GPT-5-nano on medium mazes, the difference between 80.0% and 83.33% is within sampling error, and on hard mazes the confidence intervals overlap substantially. The claim of significant reliability improvement is therefore not supported by the reported point estimates and intervals.","section":"Table 1"},{"comment":"The baseline comparison is not well controlled. Section 4 states that experiments use n=2 execution agents, but the 'Solo' rows in Table 1 appear to be single-agent runs; if so, the reported factor-of-3 improvement conflates adding an agent with adding orchestration. In addition, the random-walk floor baseline is omitted on hard mazes, and no comparison is made to a standard multi-agent ensemble without FE feedback or to classical planners (e.g., A*, BFS) that solve these mazes optimally, so the absolute improvement claim is not anchored.","section":"Section 4 and Table 1"}],"minor_comments":[{"comment":"The increment statement t←t−1 should be t←t+1, since the loop counter increases with each iteration.","section":"Algorithm 1, line 27"},{"comment":"Decimal commas and decimal points are used inconsistently (e.g., '3,03', '76,67' versus '72.22'); unify to a single convention.","section":"Abstract and Section 1"},{"comment":"The sentence 'define define interaction pathways' contains a duplicated word; also, the reference to 'Figure 8' in Section 3.1 appears to point to the wrong appendix figure, since Figure 8 in Appendix A.6.5 is the confidence-interval convergence plot, not the update sequence.","section":"Section 3.1"},{"comment":"The term 'solo agent ensembles' is confusing: if 'solo' means a single agent, the phrase 'ensembles' is misleading; if it means multiple agents without orchestration, the baseline description should state that clearly.","section":"Section 2 and Section 4"}],"recommendation":"reject","confidential_remarks":"The central formal mechanism is not operational as written: the sign convention in Eq. (2) makes the high-epistemic-drive branches unreachable, the thresholds are tuned on the same maze difficulties used for evaluation, and the orchestration component fails to improve performance in several reported configurations. Correcting these issues would require redefining the VFE objective and rerunning all experiments, which is beyond a routine revision. The appendix material is useful as prompt engineering documentation, but the paper's main claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this paper has a load-bearing flaw that the authors missed: the active-inference machinery is inert as written. Eq. (2) defines U as the negative Shannon entropy of token probabilities, so U ≤ 0 at every step. Yet the performance policies in Section 3.2 assign agents to high-epistemic-drive categories only when U > 0.6. That branch is unreachable. All agents always fall into low-drive categories, and the dynamic weight modulation (Eq. 7) can only respond to the accuracy-cost term. The claimed active-inference optimization loop never actually runs.\n\nWhat the paper does well: the cell architecture with planning/execution/orchestration nodes is clearly described, the prompt and tool scaffolding (dead-end marking, BFS backtracking, orchestrator JSON contract) is genuinely useful engineering, and the authors are honest about the adverse finding that orchestration hurts GPT-4.1-nano on hard mazes (84.6% down to 71.9%). That deserves credit.\n\nBut the soft spots are substantial. The thresholds ϑ1=0.6, ϑ2=0.4 were grid-searched on the same medium/hard maze difficulties used in the evaluation (Appendix A.5), so the reported gains are partly fitting. No code or data are shipped, despite a vague promise of a 'supplementary repository.' Baselines are thin: no comparison to other orchestration methods, no random-walk on hard, and the confidence intervals overlap for the key comparisons. And the experiment uses only two agents on synthetic mazes, so the 'scalability' claim is unsupported.\n\nThe net effect is that the reported improvements probably come from the prompt/tool scaffolding rather than from active inference. That is not nothing—the paper shows a lightweight LLM team with good prompts and a central coordinator can solve medium mazes—but it undercuts the paper's central claim of principled free-energy optimization.\n\nI'd send this to review, not desk-reject. The sign flaw is fixable (either redefine U as positive information gain or change the thresholds), the methodology can be tightened with pre-registered thresholds and released code, and the question of cheap coordination for LLM teams is worth answering. The authors need to redo the evaluation if they fix the mechanism. Who benefits: researchers building LLM multi-agent systems and active-inference practitioners looking for cautionary examples. My verdict stands: serious referee time, but substantial revision expected.","headline":"The active-inference engine in this paper cannot run as written—the epistemic term is always negative—but the engineering scaffolding and honest negative result make it worth a referee's time.","tokens_in":17889,"tokens_out":2702,"would_cite":false,"duration_ms":22880,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Orchestrator claims that a lightweight multi-agent LLM system coordinated by an active-inference free-energy signal solves long-horizon mazes at up to 100% success on medium 18x18 mazes and 76.67% on hard 25x25 mazes, outperforming solo…","keywords":["active inference","multi-agent systems","large language models","variational free energy","long-horizon tasks","maze solving","partial observability","orchestration"],"falsifier":"Run the same maze set with the free-energy signal replaced by random or constant weights, keeping everything else identical; if success rates stay at 100% and 76.67%, the free-energy computation is not doing the work. Alternatively, generate a very-hard 30x30 maze set (the paper defines one but does not test it) and apply the same thresholds; if success collapses, the thresholds were overfit to the tested difficulties.","tokens_in":16898,"feed_emoji":"🧩","tokens_out":8731,"duration_ms":65812,"temperature":0.7,"pith_summary":"This paper claims that a cheap multi-agent LLM system can be coordinated through active inference, monitoring each agent's variational free energy, the balance between information gain and behavioral cost, rather than relying on larger models or hand-coded plans. On procedurally generated mazes of increasing size, the Orchestrator framework reports near-perfect success on medium 18x18 mazes (100% for one small model, 83% for another) and up to 76.67% on hard 25x25 mazes, versus 0–30% for solo agents. The claim matters because it suggests long-horizon, partially observable coordination problems do not require expensive reasoning models: a lightweight monitoring loop plus two small LLMs may suffice. The paper also reports that the free-energy benchmark alone, without an orchestrator node, already lifts success rates substantially, and that adding orchestration can hurt on the hardest mazes because of reasoning overhead.","feed_headline":"Active-inference orchestration lifts cheap LLM teams to 100%","feed_subtitle":"Adding a free-energy monitor and orchestrator node lifts cheap LLM teams to 100% on medium mazes and 77% on hard ones.","key_machinery":"The load-bearing object is the operational variational free energy $F_n(t,k) = U_{\\text{epistemic}}(n,t,k) - C_{\\text{accuracy}}(n,t,k)$, where $U_{\\text{epistemic}} = -H[S_{n,t,k} \\mid S_{n-1,t-1,k-1}]$ is the negative Shannon entropy of the agent's message tokens (interpreted as realized information gain) and $C_{\\text{accuracy}}$ is an equally weighted sum of five risk components: movement efficiency, exploration efficiency, backtracking patterns, dead-end recognition, and oscillation avoidance. Thresholds $\\vartheta_1 = 0.6$ and $\\vartheta_2 = 0.4$ divide agents into four performance categories, and each category triggers updates to exploration, exploitation, coordination, and backtracking weights; an orchestrator node additionally injects corrective prompts. This machinery is what converts active inference from a principle into a control signal.","core_discovery":"The central discovery is that a multi-agent system's global task performance can be optimized by giving each LLM agent a scalar variational free energy score at every step, computed as the negative Shannon entropy of its token outputs minus a weighted penalty for inefficient behaviors, and then using thresholds on that score to adjust exploration, exploitation, coordination, and backtracking weights. With two small LLMs as execution agents and one orchestrator node acting as shared memory and prompt injector, the system solves 18x18 mazes of medium difficulty at 100% success (25 runs) and 25x25 hard mazes at up to 76.67% success, outperforming solo-agent baselines by an average factor of 3.03 on medium mazes. The paper interprets these results as evidence that active-inference benchmarking mitigates partial observability and helps agents escape local minima.","pith_inferences":["If the free-energy proxy generalizes beyond mazes, the same two-signal control loop (information gain minus behavioral cost) could be dropped into other partially observable long-horizon domains, such as warehouse navigation or multi-agent search, without retraining the LLMs.","The paper tunes its two thresholds on medium and hard mazes by grid search; a held-out difficulty or a randomized threshold sweep would reveal how much of the gain is active inference versus tuning.","Only two execution agents are reported as a balanced trade-off; testing one or three to four agents would clarify whether the orchestrator's benefit scales with team size or saturates.","Because the epistemic term is token entropy, not map-based information gain, one could test whether replacing it with a direct measure of newly discovered maze cells changes performance; if not, the LLM's token uncertainty is an unnecessary middleman."],"forward_implications":["Free-energy benchmarking alone lifts medium-maze success from about 30% (or 0%) for solo small models to 72–80%, so cheap introspection can substitute for a larger model.","Adding an orchestrator node pushes medium-maze success to 100% for one small model and 83% for another, but on hard mazes it can reduce success (e.g., from 84.6% to 71.9% for one model), implying orchestration overhead has a cost at the edge of the model's planning horizon.","The framework achieves these results with small, fast-inference LLMs, suggesting production deployments on strict API budgets are feasible.","The reported factor of 3.03 improvement over solo ensembles on medium mazes is the headline quantitative claim of the paper."],"supporting_citations":[{"why":"Supplies the active-inference / variational-free-energy formalism that the paper operationalizes.","marker":"[29]"},{"why":"Provides the maze benchmark generator and complexity metrics used for all experiments.","marker":"[36]"},{"why":"Justifies treating the unknown posterior as uniform via the maximum entropy principle in the VFE derivation.","marker":"[57]"},{"why":"Supports the use of active-inference agents in delayed and long-horizon environments.","marker":"[33]"},{"why":"Provides the reflective multi-agent collaboration baseline that Orchestrator extends.","marker":"[30]"},{"why":"Supplies the evolving orchestration approach that motivates the dynamic graph structure.","marker":"[23]"}],"fun_headline_variants":["Free-energy metric guides cheap LLM teams to 100% on medium mazes","Active-inference scores take LLM teams to 100% on medium mazes","Orchestrator's free-energy monitor lifts cheap LLMs to 100% on medium mazes","LLM teams hit 100% on medium mazes via free-energy orchestration","Active inference helps cheap LLM teams crack 100% on medium mazes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire result depends on the paper's formula for a free-energy score (how surprising the agent's outputs are minus how wasteful its moves are) actually measuring good exploration; if that formula is wrong or the thresholds were picked on the test mazes, the gains are just tuned heuristics.","fun_headline_variants_meta":{"raw":{"variants":["Free-energy metric guides cheap LLM teams to 100% on medium mazes","Active-inference scores take LLM teams to 100% on medium mazes","Orchestrator's free-energy monitor lifts cheap LLMs to 100% on medium mazes","LLM teams hit 100% on medium mazes via free-energy orchestration","Active inference helps cheap LLM teams crack 100% on medium mazes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001723,"raw_usage":{"total_tokens":6764,"prompt_tokens":842,"completion_tokens":5922,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":5813}},"tokens_in":458,"tokens_out":5922,"duration_ms":37031,"temperature":1.0,"reasoning_tokens":5813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:21:23.253470+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same maze set with the free-energy signal replaced by random or constant weights, keeping everything else identical; if success rates stay at 100% and 76.67%, the free-energy computation is not doing the work. Alternatively, generate a very-hard 30x30 maze set (the paper defines one but does not test it) and apply the same thresholds; if success collapses, the thresholds were overfit to the tested difficulties.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the active-inference / variational-free-energy formalism that the paper operationalizes."},{"cited_title":"AMaze: An Intuitive Benchmark Generator for Fast Prototyping of Generalizable Agents","cited_arxiv_id":null,"evidence_quote":"Provides the maze benchmark generator and complexity metrics used for all experiments."},{"cited_title":"LookAround","cited_arxiv_id":null,"evidence_quote":"Justifies treating the unknown posterior as uniform via the maximum entropy principle in the VFE derivation."},{"cited_title":"Reflective Multi-Agent Collaboration based on Large Language Models","cited_arxiv_id":null,"evidence_quote":"Provides the reflective multi-agent collaboration baseline that Orchestrator extends."}],"review_version":2}