Hallucination as a Feature, not a Defect: Evaluating a multi-agent architecture to transform speculative language-model outputs into testable scientific hypotheses
不消灭AI的'胡编乱造',而是把它关进笼子里逼它产出可检验的研究假设
一名西班牙高中生(相当于高一)用Rust搭建了一个多智能体系统,让一个AI自由发散地生成想法,另一个AI用网络搜索去核实并批评它,中间再加一个过滤器减少重复和噪音。这项工作想验证,把大语言模型的'幻觉'(一本正经编造内容的倾向)加以控制,是否能把它变成产生科学假设的有用工具,实验在物理和社会科学类问题上进行。结果显示,完整系统明显强于直接问模型,但并没有全面超过让模型自己反思修改这种更简单的方法。
他们做了什么
- 大语言模型通常被训练来抑制'幻觉'(一本正经地编造不实内容),这项工作想看能否反过来,把这种倾向控制住,用来生成研究假设。
- 架构让一个高随机性的生成智能体自由产生想法,另一个依靠网络搜索的评估智能体对其提出批评,中间用一个低随机性的语义过滤器减少重复和噪音,作者称这个循环为'认识论摩擦'。
- 研究在海水淡化、议会僵局等问题上比较了六种条件:直接提问、自我反思、完整系统,以及去掉语义过滤器、去掉搜索核实、去掉侧向视角这三种消融实验。
- 直接提问在大多数指标上表现最弱,但完整系统并没有全面优于简单的自我反思;去掉侧向视角能提高可行性,却系统性地降低了组合稀有度。
- 完整系统的主要优势体现在受到较强物理、实证或制度约束的问题上,而在更偏概念、协商性的问题上,单纯的自我反思往往就足够了。

| Problem | Condition | Generated hypothesis | Qualitative reading |
|---|---|---|---|
| Desalination | direct_llm | Electromagnetic pumping of seawater in the solid state. Proposes using crossed electric and magnetic fields (MHD) or traveling magnetic fields to move seawater and brine without mechanical parts. | The idea is exotic, but it does not directly solve salt separation: it replaces mechanical pumping with electromagnetic pumping. It was rated low feasibility (F=1) with null structural validity. |
| Desalination | self_reflection | Resonant acoustic freeze desalination. Combines bipolar sub-nanosecond electrical pulses to weaken hydration layers with acoustic ice nucleation via ultrasound. | Self-reflection produces a much more polished and metrically strong hypothesis (F=3, Q=1.000), but it retains a significant dependence on active fields and fine acoustic control. It improves the argumentative envelope without fully shifting the mechanism toward an external constraint. |
| Desalination | full_system | Cryo-flotation desalination with cyclopentane clathrates. Uses cyclopentane to form clathrates at 7∘C, geometrically exclude salt, separate by buoyancy, and couple the cycle to an industrial cryogenic source such as LNG regasification. | The critique discarded MHD, cavitation, and active fields due to parasitic heating or dielectric breakdown. The result retains the useful core: salt exclusion by clathrates, explicit management of latent heat, and a passive separation mechanism. |
| Parliamentary deadlock | direct_llm | Meritocratic delegated voting with zero-knowledge proofs. Assigns different voting weight to deputies according to a merit score computed by an external oracle and applies a cryptographic penalty to party discipline. | The proposal introduces a flashy but democratically fragile solution: an unelected oracle alters the voting weight of elected representatives. It was rated low feasibility (F=1) with null structural validity. |
| Parliamentary deadlock | self_reflection | Randomized evidence-prediction tournaments. Randomly distributes deputies into teams that predict the future validity of legislative evidence and accumulate reputation for accuracy. | A pragmatic, high-quality output (F=3, Q=1.000): it could work as an advisory tool in committees. However, it mainly addresses the epistemic quality of deliberation, not the coercive core of party discipline. |
| Parliamentary deadlock | full_system | Electoral epistemic deposit with adversarial resolution. Deputies dissent by betting their future position on electoral lists against exogenous legislative indicators, verified by a multi-party adversarial jury. | The critical pressure discards secret voting, vote weighting, and unconstitutional mechanisms. The final proposal preserves formal equality of the vote and shifts the incentive to the internal party cost, making epistemological friction more tangible. |

| Condition | Structural validity | Feasible rate | Originality | Comb. rarity | Diversity | Collapse |
|---|---|---|---|---|---|---|
| direct_llm | 0.411 | 0.678 | 0.389 | 0.491 | 0.776 | 0.417 |
| self_reflection | 0.656 | 0.950 | 0.622 | 0.830 | 0.790 | 0.392 |
| no_agent4 | 0.500 | 0.967 | 0.607 | 0.757 | 0.733 | 0.458 |
| no_search | 0.411 | 0.936 | 0.651 | 0.745 | 0.821 | 0.403 |
| no_lateral_lenses | 0.561 | 1.000 | 0.647 | 0.670 | 0.818 | 0.432 |
| full_system | 0.444 | 0.944 | 0.652 | 0.848 | 0.766 | 0.405 |

| Condition | Structural validity | Feasible rate | Originality | Comb. rarity | FCCY index |
|---|---|---|---|---|---|
| self_reflection | 0.656 | 0.950 | 0.622 | 0.830 | 0.321 |
| no_lateral_lenses | 0.561 | 1.000 | 0.647 | 0.670 | 0.243 |
| full_system | 0.444 | 0.944 | 0.652 | 0.848 | 0.232 |
| no_agent4 | 0.500 | 0.967 | 0.607 | 0.757 | 0.222 |
| no_search | 0.411 | 0.936 | 0.651 | 0.745 | 0.187 |
| direct_llm | 0.411 | 0.678 | 0.389 | 0.491 | 0.053 |

| Condition | Fluency | Quality | Usefulness | Creativity | SVHN rate | VCN | Valid rare |
|---|---|---|---|---|---|---|---|
| direct_llm | 1.100 | 0.736 | 0.802 | 0.567 | 0.144 | 0.174 | 0.133 |
| self_reflection | 1.567 | 0.793 | 0.844 | 0.713 | 0.656 | 0.534 | 0.600 |
| no_agent4 | 1.500 | 0.778 | 0.833 | 0.696 | 0.467 | 0.343 | 0.367 |
| no_search | 2.000 | 0.759 | 0.819 | 0.726 | 0.411 | 0.278 | 0.367 |
| no_lateral_lenses | 1.533 | 0.793 | 0.845 | 0.728 | 0.561 | 0.423 | 0.467 |
| full_system | 1.767 | 0.770 | 0.827 | 0.719 | 0.444 | 0.381 | 0.467 |

| Condition | Expected | Completed | Failed | Compl. rate | API issues | API limit | Other technical |
|---|---|---|---|---|---|---|---|
| direct_llm | 30 | 30 | 0 | 1.000 | 0.000 | 0.000 | 0.000 |
| self_reflection | 30 | 30 | 0 | 1.000 | 0.000 | 0.000 | 0.000 |
| no_agent4 | 30 | 30 | 0 | 1.000 | 0.000 | 0.000 | 0.000 |
| no_search | 30 | 30 | 0 | 1.000 | 0.000 | 0.000 | 0.000 |
| no_lateral_lenses | 30 | 30 | 0 | 1.000 | 0.000 | 0.000 | 0.000 |
| full_system | 30 | 30 | 0 | 1.000 | 0.000 | 0.000 | 0.000 |

| Comparator | Metric | Blocks | 𝚫 mean | Win rate | Cohen dz | 95% CI |
|---|---|---|---|---|---|---|
| direct_llm | Feasibility | 3 | 0.139 | 1.000 | 1.981 | [0.075; 0.214] |
| direct_llm | Originality | 3 | 0.263 | 1.000 | 7.896 | [0.235; 0.300] |
| direct_llm | Comb. rarity | 3 | 0.357 | 1.000 | 1.265 | [0.068; 0.631] |
| direct_llm | Diversity | 2 | 0.006 | 0.500 | 0.064 | [-0.056; 0.067] |
| direct_llm | Collapse | 2 | -0.031 | 0.500 | -0.193 | [-0.146; 0.083] |
| direct_llm | Creativity | 3 | 0.153 | 1.000 | 3.774 | [0.108; 0.187] |
| self_reflection | Feasibility | 3 | 0.021 | 0.667 | 0.346 | [-0.044; 0.078] |
| self_reflection | Originality | 3 | 0.031 | 1.000 | 2.033 | [0.013; 0.041] |
| self_reflection | Comb. rarity | 3 | 0.019 | 0.667 | 0.231 | [-0.073; 0.078] |
| self_reflection | Diversity | 3 | -0.005 | 0.667 | -0.113 | [-0.060; 0.025] |
| self_reflection | Collapse | 3 | -0.021 | 0.333 | -0.375 | [-0.071; 0.038] |
| self_reflection | Creativity | 3 | 0.006 | 0.667 | 0.283 | [-0.013; 0.029] |
| no_agent4 | Feasibility | 3 | 0.011 | 0.667 | 0.296 | [-0.028; 0.047] |
| no_agent4 | Originality | 3 | 0.045 | 1.000 | 0.633 | [0.002; 0.128] |
| no_agent4 | Comb. rarity | 3 | 0.091 | 0.667 | 0.605 | [-0.017; 0.264] |
| no_agent4 | Diversity | 3 | 0.065 | 1.000 | 0.777 | [0.000; 0.159] |
| no_agent4 | Collapse | 3 | -0.046 | 0.000 | -2.639 | [-0.062; -0.028] |
| no_agent4 | Creativity | 3 | 0.023 | 1.000 | 0.784 | [0.003; 0.058] |
| no_search | Feasibility | 3 | -0.011 | 0.333 | -0.158 | [-0.086; 0.053] |
| no_search | Originality | 3 | 0.001 | 0.667 | 0.036 | [-0.037; 0.028] |
| no_search | Comb. rarity | 3 | 0.103 | 1.000 | 1.014 | [0.037; 0.220] |
| no_search | Diversity | 3 | -0.031 | 0.333 | -0.291 | [-0.123; 0.087] |
| no_search | Collapse | 3 | -0.002 | 0.667 | -0.042 | [-0.052; 0.026] |
| no_search | Creativity | 3 | -0.007 | 0.333 | -0.189 | [-0.036; 0.033] |
| no_lateral_lenses | Feasibility | 3 | -0.049 | 0.000 | —* | [-0.050; -0.047] |
| no_lateral_lenses | Originality | 3 | 0.006 | 0.333 | 0.090 | [-0.040; 0.079] |
| no_lateral_lenses | Comb. rarity | 3 | 0.178 | 1.000 | 3.924 | [0.147; 0.230] |
| no_lateral_lenses | Diversity | 3 | -0.027 | 0.333 | -0.428 | [-0.085; 0.042] |
| no_lateral_lenses | Collapse | 3 | -0.015 | 0.333 | -0.283 | [-0.062; 0.042] |
| no_lateral_lenses | Creativity | 3 | -0.009 | 0.333 | -0.223 | [-0.039; 0.034] |
| Condition | Struct. valid. | Feasibility | Originality | Comb. rarity | Diversity | Collapse | Creativity |
|---|---|---|---|---|---|---|---|
| direct_llm | 4.667 | 6.000 | 6.000 | 5.667 | 4.000 | 3.500 | 6.000 |
| self_reflection | 1.667 | 2.333 | 4.000 | 2.000 | 3.000 | 3.667 | 3.333 |
| no_agent4 | 3.000 | 3.333 | 4.000 | 3.000 | 4.333 | 4.667 | 4.333 |
| no_search | 5.000 | 3.667 | 3.000 | 4.000 | 3.000 | 2.333 | 2.667 |
| no_lateral_lenses | 2.667 | 2.333 | 2.000 | 4.667 | 3.000 | 3.333 | 2.000 |
| full_system | 4.000 | 3.333 | 2.000 | 1.667 | 3.000 | 2.667 | 2.667 |
| Problem | Condition | Structural validity | Feasible rate | Originality | Comb. rarity | SVHN rate | Quality |
|---|---|---|---|---|---|---|---|
| Desalination | direct_llm | 0.367 | 0.617 | 0.350 | 0.651 | 0.067 | 0.725 |
| Desalination | self_reflection | 0.750 | 0.925 | 0.601 | 0.869 | 0.750 | 0.794 |
| Desalination | no_agent4 | 0.483 | 0.975 | 0.623 | 0.866 | 0.433 | 0.780 |
| Desalination | no_search | 0.483 | 0.904 | 0.632 | 0.743 | 0.483 | 0.756 |
| Desalination | no_lateral_lenses | 0.700 | 1.000 | 0.658 | 0.678 | 0.700 | 0.812 |
| Desalination | full_system | 0.567 | 0.933 | 0.627 | 0.871 | 0.567 | 0.784 |
| Parliament | direct_llm | 0.500 | 0.800 | 0.468 | 0.171 | 0.300 | 0.758 |
| Parliament | self_reflection | 0.467 | 1.000 | 0.663 | 0.751 | 0.467 | 0.791 |
| Parliament | no_agent4 | 0.533 | 0.950 | 0.575 | 0.538 | 0.533 | 0.773 |
| Parliament | no_search | 0.267 | 1.000 | 0.690 | 0.750 | 0.267 | 0.765 |
| Parliament | no_lateral_lenses | 0.283 | 1.000 | 0.625 | 0.656 | 0.283 | 0.756 |
| Parliament | full_system | 0.200 | 0.967 | 0.704 | 0.802 | 0.200 | 0.741 |
| Use | Condition | Hypothesis | Qualitative reading | Provenance and metrics |
|---|---|---|---|---|
| Main | direct_llm | Electromagnetic pumping of seawater in the solid state | Striking but misdirected idea: it replaces mechanical pumping with MHD without directly solving ionic separation. It serves as an example of attractive speculation with low physical constraint. | Run 0065, id 3. Desalination, seed 43, repetition 05. F=1, Q=0.528. |
| Main | self_reflection | Resonant acoustic freeze desalination | Self-reflection greatly improves the final form and reaches a high score, but it retains a strong dependence on sub-nanosecond PEF and precise acoustic nucleation. | Run 0073, id 2. Desalination, seed 43, repetition 03. F=3, Q=1.000. |
| Main | full_system | Cryo-flotation desalination with cyclopentane clathrates | External critique shifts the design toward a more passive separation mechanism: cyclopentane clathrates, buoyancy, and explicit management of latent heat. | Run 0117, id 1. Desalination, seed 43, repetition 07. F=3, Q=0.917. |
| Main | direct_llm | Meritocratic delegated voting with zero-knowledge proofs | Introduces a merit oracle that alters parliamentary voting weight. The proposal is sophisticated on the surface, but politically fragile and structurally invalid. | Run 0129, id 1. Parliament, seed 42, repetition 09. F=1, Q=0.694. |
| Main | self_reflection | Randomized evidence-prediction tournaments | Strong example of self-reflection: it proposes an implementable advisory mechanism to improve evidence evaluation, although it does not directly address party coercion. | Run 0137, id 3. Parliament, seed 42, repetition 07. F=3, Q=1.000. |
| Main | full_system | Electoral epistemic deposit with adversarial resolution | The final output avoids weighting votes or using parliamentary secrecy and shifts the incentive to the internal party cost, preserving formal equality of deputies. | Run 0171, id 1. Parliament, seed 42, repetition 01. F=2, Q=0.889. |
| Appendix | no_search | Supersonic ballistic ion classifier by MHD | Without external search, a spectacular MHD solution reappears: fields >10T, oceanic pressure, and Lorentz deflection. It is creative, but very vulnerable to turbulent mixing and physical scale. | Run 0098, id 1. Desalination, seed 43, repetition 08. F=2, Q=0.806. |
| Appendix | no_agent4 | Microfluidic polarization of ionic concentration | The absence of the last agent does not prevent producing a solid hypothesis: it reduces macroscopic MHD to microfluidic electrokinetics, a physically more defensible route. | Run 0022, id 3. Desalination, seed 42, repetition 02. F=3, Q=0.917. |
| Appendix | no_lateral_lenses | Electrochemical ion sponge based on Prussian blue analogues | Removing lateral lenses yields a less exotic solution, closer to existing electrochemical intercalation technologies. It gains plausibility but loses part of the combinatorial leap. | Run 0042, id 1. Desalination, seed 42, repetition 02. F=3, Q=0.917. |
| Appendix | no_search | Quadratic time market to control the agenda | Proposes a procedural economy where the cost of hoarding parliamentary time grows quadratically. It is elegant, although it depends on an institutional formalization not externally contrasted. | Run 0155, id 1. Parliament, seed 42, repetition 05. F=3, Q=0.917. |
| Appendix | no_agent4 | Zero-knowledge activation of Article 85.1 | A particularly valuable example: it uses a local legal anchor, the 70-signature threshold of Article 85.1, and applies cryptography only to protect internal coordination. | Run 0150, id 1. Parliament, seed 42, repetition 10. F=3, Q=0.917. |
| Appendix | no_lateral_lenses | Consensus-weighted procedural dividend | A sober and probably implementable solution: it rewards multi-party consensus with a procedural fast track. It shows that removing laterality can increase institutional clarity. | Run 0167, id 2. Parliament, seed 42, repetition 07. F=3, Q=0.917. |
| Problema | Condición | Hipótesis generada | Lectura cualitativa |
|---|---|---|---|
| Desalinización | direct_llm | Bombeo electromagnético de agua de mar en estado sólido. Propone usar campos eléctricos y magnéticos cruzados (MHD) o campos magnéticos viajeros para mover agua de mar y salmuera sin piezas mecánicas. | La idea es exótica, pero no resuelve directamente la separación de sal: sustituye el bombeo mecánico por bombeo electromagnético. Fue evaluada con baja factibilidad (F=1) y validez estructural nula. |
| Desalinización | self_reflection | Desalinización por congelación acústica resonante. Combina pulsos eléctricos bipolares subnanosegundo para debilitar capas de hidratación con nucleación acústica de hielo mediante ultrasonidos. | La autorreflexión produce una hipótesis mucho más pulida y métricamente fuerte (F=3, Q=1,000), pero conserva una dependencia importante de campos activos y control acústico fino. Mejora el envoltorio argumental sin desplazar completamente el mecanismo hacia una restricción externa. |
| Desalinización | full_system | Desalinización crioflotante con clatratos de ciclopentano. Usa ciclopentano para formar clatratos a 7∘C, excluir geométricamente la sal, separar por flotabilidad y acoplar el ciclo a una fuente criogénica industrial como regasificación de GNL. | La crítica descartó MHD, cavitación y campos activos por calentamiento parásito o ruptura dieléctrica. El resultado conserva el núcleo útil: exclusión salina por clatratos, gestión explícita del calor latente y mecanismo pasivo de separación. |
| Bloqueo parlamentario | direct_llm | Voto delegado meritocrático con pruebas de conocimiento cero. Asigna distinto peso de voto a los diputados según una puntuación de mérito calculada por un oráculo externo y aplica una penalización criptográfica a la disciplina de partido. | La propuesta introduce una solución vistosa, pero democráticamente frágil: un oráculo no electo altera el peso del voto de representantes electos. Fue evaluada con baja factibilidad (F=1) y validez estructural nula. |
| Bloqueo parlamentario | self_reflection | Torneos aleatorizados de predicción de evidencia. Distribuye aleatoriamente a diputados en equipos que predicen la validez futura de la evidencia legislativa y acumulan reputación por acierto. | Es una salida pragmática y de alta calidad (F=3, Q=1,000): podría funcionar como herramienta asesora en comités. Sin embargo, ataca sobre todo la calidad epistémica de la deliberación, no el núcleo coercitivo de la disciplina de partido. |
| Bloqueo parlamentario | full_system | Depósito epistémico electoral con resolución adversarial. Los diputados disienten apostando su futura posición en listas electorales contra indicadores legislativos exógenos, verificados por un jurado adversarial multipartidista. | La presión crítica descarta voto secreto, ponderación de votos y mecanismos inconstitucionales. La propuesta final mantiene igualdad formal del voto y desplaza el incentivo al coste interno de partido, haciendo más tangible la fricción epistemológica. |
| Condición | Validez estructural | Tasa factible | Originalidad | Rareza comb. | Diversidad | Colapso |
|---|---|---|---|---|---|---|
| direct_llm | 0,411 | 0,678 | 0,389 | 0,491 | 0,776 | 0,417 |
| self_reflection | 0,656 | 0,950 | 0,622 | 0,830 | 0,790 | 0,392 |
| no_agent4 | 0,500 | 0,967 | 0,607 | 0,757 | 0,733 | 0,458 |
| no_search | 0,411 | 0,936 | 0,651 | 0,745 | 0,821 | 0,403 |
| no_lateral_lenses | 0,561 | 1,000 | 0,647 | 0,670 | 0,818 | 0,432 |
| full_system | 0,444 | 0,944 | 0,652 | 0,848 | 0,766 | 0,405 |
| Condición | Validez estructural | Tasa factible | Originalidad | Rareza comb. | Índice FCCY |
|---|---|---|---|---|---|
| self_reflection | 0,656 | 0,950 | 0,622 | 0,830 | 0,321 |
| no_lateral_lenses | 0,561 | 1,000 | 0,647 | 0,670 | 0,243 |
| full_system | 0,444 | 0,944 | 0,652 | 0,848 | 0,232 |
| no_agent4 | 0,500 | 0,967 | 0,607 | 0,757 | 0,222 |
| no_search | 0,411 | 0,936 | 0,651 | 0,745 | 0,187 |
| direct_llm | 0,411 | 0,678 | 0,389 | 0,491 | 0,053 |
| Condición | Fluidez | Calidad | Utilidad | Creatividad | Tasa SVHN | VCN | Válidas raras |
|---|---|---|---|---|---|---|---|
| direct_llm | 1,100 | 0,736 | 0,802 | 0,567 | 0,144 | 0,174 | 0,133 |
| self_reflection | 1,567 | 0,793 | 0,844 | 0,713 | 0,656 | 0,534 | 0,600 |
| no_agent4 | 1,500 | 0,778 | 0,833 | 0,696 | 0,467 | 0,343 | 0,367 |
| no_search | 2,000 | 0,759 | 0,819 | 0,726 | 0,411 | 0,278 | 0,367 |
| no_lateral_lenses | 1,533 | 0,793 | 0,845 | 0,728 | 0,561 | 0,423 | 0,467 |
| full_system | 1,767 | 0,770 | 0,827 | 0,719 | 0,444 | 0,381 | 0,467 |
| Condición | Esperadas | Completadas | Fallidas | Tasa complet. | Problemas API | Límite API | Otros técnicos |
|---|---|---|---|---|---|---|---|
| direct_llm | 30 | 30 | 0 | 1,000 | 0,000 | 0,000 | 0,000 |
| self_reflection | 30 | 30 | 0 | 1,000 | 0,000 | 0,000 | 0,000 |
| no_agent4 | 30 | 30 | 0 | 1,000 | 0,000 | 0,000 | 0,000 |
| no_search | 30 | 30 | 0 | 1,000 | 0,000 | 0,000 | 0,000 |
| no_lateral_lenses | 30 | 30 | 0 | 1,000 | 0,000 | 0,000 | 0,000 |
| full_system | 30 | 30 | 0 | 1,000 | 0,000 | 0,000 | 0,000 |
| Comparador | Métrica | Bloques | 𝚫 medio | Tasa victoria | Cohen dz | IC 95 % |
|---|---|---|---|---|---|---|
| direct_llm | Factibilidad | 3 | 0,139 | 1,000 | 1,981 | [0,075; 0,214] |
| direct_llm | Originalidad | 3 | 0,263 | 1,000 | 7,896 | [0,235; 0,300] |
| direct_llm | Rareza comb. | 3 | 0,357 | 1,000 | 1,265 | [0,068; 0,631] |
| direct_llm | Diversidad | 2 | 0,006 | 0,500 | 0,064 | [-0,056; 0,067] |
| direct_llm | Colapso | 2 | -0,031 | 0,500 | -0,193 | [-0,146; 0,083] |
| direct_llm | Creatividad | 3 | 0,153 | 1,000 | 3,774 | [0,108; 0,187] |
| self_reflection | Factibilidad | 3 | 0,021 | 0,667 | 0,346 | [-0,044; 0,078] |
| self_reflection | Originalidad | 3 | 0,031 | 1,000 | 2,033 | [0,013; 0,041] |
| self_reflection | Rareza comb. | 3 | 0,019 | 0,667 | 0,231 | [-0,073; 0,078] |
| self_reflection | Diversidad | 3 | -0,005 | 0,667 | -0,113 | [-0,060; 0,025] |
| self_reflection | Colapso | 3 | -0,021 | 0,333 | -0,375 | [-0,071; 0,038] |
| self_reflection | Creatividad | 3 | 0,006 | 0,667 | 0,283 | [-0,013; 0,029] |
| no_agent4 | Factibilidad | 3 | 0,011 | 0,667 | 0,296 | [-0,028; 0,047] |
| no_agent4 | Originalidad | 3 | 0,045 | 1,000 | 0,633 | [0,002; 0,128] |
| no_agent4 | Rareza comb. | 3 | 0,091 | 0,667 | 0,605 | [-0,017; 0,264] |
| no_agent4 | Diversidad | 3 | 0,065 | 1,000 | 0,777 | [0,000; 0,159] |
| no_agent4 | Colapso | 3 | -0,046 | 0,000 | -2,639 | [-0,062; -0,028] |
| no_agent4 | Creatividad | 3 | 0,023 | 1,000 | 0,784 | [0,003; 0,058] |
| no_search | Factibilidad | 3 | -0,011 | 0,333 | -0,158 | [-0,086; 0,053] |
| no_search | Originalidad | 3 | 0,001 | 0,667 | 0,036 | [-0,037; 0,028] |
| no_search | Rareza comb. | 3 | 0,103 | 1,000 | 1,014 | [0,037; 0,220] |
| no_search | Diversidad | 3 | -0,031 | 0,333 | -0,291 | [-0,123; 0,087] |
| no_search | Colapso | 3 | -0,002 | 0,667 | -0,042 | [-0,052; 0,026] |
| no_search | Creatividad | 3 | -0,007 | 0,333 | -0,189 | [-0,036; 0,033] |
| no_lateral_lenses | Factibilidad | 3 | -0,049 | 0,000 | —* | [-0,050; -0,047] |
| no_lateral_lenses | Originalidad | 3 | 0,006 | 0,333 | 0,090 | [-0,040; 0,079] |
| no_lateral_lenses | Rareza comb. | 3 | 0,178 | 1,000 | 3,924 | [0,147; 0,230] |
| no_lateral_lenses | Diversidad | 3 | -0,027 | 0,333 | -0,428 | [-0,085; 0,042] |
| no_lateral_lenses | Colapso | 3 | -0,015 | 0,333 | -0,283 | [-0,062; 0,042] |
| no_lateral_lenses | Creatividad | 3 | -0,009 | 0,333 | -0,223 | [-0,039; 0,034] |
| Condición | Validez estr. | Factibilidad | Originalidad | Rareza comb. | Diversidad | Colapso | Creatividad |
|---|---|---|---|---|---|---|---|
| direct_llm | 4,667 | 6,000 | 6,000 | 5,667 | 4,000 | 3,500 | 6,000 |
| self_reflection | 1,667 | 2,333 | 4,000 | 2,000 | 3,000 | 3,667 | 3,333 |
| no_agent4 | 3,000 | 3,333 | 4,000 | 3,000 | 4,333 | 4,667 | 4,333 |
| no_search | 5,000 | 3,667 | 3,000 | 4,000 | 3,000 | 2,333 | 2,667 |
| no_lateral_lenses | 2,667 | 2,333 | 2,000 | 4,667 | 3,000 | 3,333 | 2,000 |
| full_system | 4,000 | 3,333 | 2,000 | 1,667 | 3,000 | 2,667 | 2,667 |
| Problema | Condición | Validez estructural | Tasa factible | Originalidad | Rareza comb. | Tasa SVHN | Calidad |
|---|---|---|---|---|---|---|---|
| Desalinización | direct_llm | 0,367 | 0,617 | 0,350 | 0,651 | 0,067 | 0,725 |
| Desalinización | self_reflection | 0,750 | 0,925 | 0,601 | 0,869 | 0,750 | 0,794 |
| Desalinización | no_agent4 | 0,483 | 0,975 | 0,623 | 0,866 | 0,433 | 0,780 |
| Desalinización | no_search | 0,483 | 0,904 | 0,632 | 0,743 | 0,483 | 0,756 |
| Desalinización | no_lateral_lenses | 0,700 | 1,000 | 0,658 | 0,678 | 0,700 | 0,812 |
| Desalinización | full_system | 0,567 | 0,933 | 0,627 | 0,871 | 0,567 | 0,784 |
| Parlamento | direct_llm | 0,500 | 0,800 | 0,468 | 0,171 | 0,300 | 0,758 |
| Parlamento | self_reflection | 0,467 | 1,000 | 0,663 | 0,751 | 0,467 | 0,791 |
| Parlamento | no_agent4 | 0,533 | 0,950 | 0,575 | 0,538 | 0,533 | 0,773 |
| Parlamento | no_search | 0,267 | 1,000 | 0,690 | 0,750 | 0,267 | 0,765 |
| Parlamento | no_lateral_lenses | 0,283 | 1,000 | 0,625 | 0,656 | 0,283 | 0,756 |
| Parlamento | full_system | 0,200 | 0,967 | 0,704 | 0,802 | 0,200 | 0,741 |
| Uso | Condición | Hipótesis | Lectura cualitativa | Procedencia y métricas |
|---|---|---|---|---|
| Principal | direct_llm | Bombeo electromagnético de agua de mar en estado sólido | Idea llamativa pero mal orientada: sustituye el bombeo mecánico por MHD sin resolver directamente la separación iónica. Sirve como ejemplo de especulación atractiva con baja constricción física. | Ejecución 0065, id 3. Desalinización, semilla 43, repetición 05. F=1, Q=0,528. |
| Principal | self_reflection | Desalinización por congelación acústica resonante | La autorreflexión mejora mucho la forma final y alcanza alta puntuación, pero conserva una dependencia fuerte de PEF subnanosegundo y nucleación acústica precisa. | Ejecución 0073, id 2. Desalinización, semilla 43, repetición 03. F=3, Q=1,000. |
| Principal | full_system | Desalinización crioflotante con clatratos de ciclopentano | La crítica externa desplaza el diseño hacia un mecanismo de separación más pasivo: clatratos de ciclopentano, flotabilidad y gestión explícita del calor latente. | Ejecución 0117, id 1. Desalinización, semilla 43, repetición 07. F=3, Q=0,917. |
| Principal | direct_llm | Voto delegado meritocrático con pruebas de conocimiento cero | Introduce un oráculo de mérito que altera el peso del voto parlamentario. La propuesta es sofisticada en superficie, pero políticamente frágil y estructuralmente inválida. | Ejecución 0129, id 1. Parlamento, semilla 42, repetición 09. F=1, Q=0,694. |
| Principal | self_reflection | Torneos aleatorizados de predicción de evidencia | Ejemplo fuerte de autorreflexión: propone un mecanismo asesor implementable para mejorar la evaluación de evidencia, aunque no ataca directamente la coerción partidista. | Ejecución 0137, id 3. Parlamento, semilla 42, repetición 07. F=3, Q=1,000. |
| Principal | full_system | Depósito epistémico electoral con resolución adversarial | La salida final evita ponderar votos o usar secreto parlamentario y desplaza el incentivo al coste interno de partido, manteniendo igualdad formal de los diputados. | Ejecución 0171, id 1. Parlamento, semilla 42, repetición 01. F=2, Q=0,889. |
| Apéndice | no_search | Clasificador balístico supersónico de iones por MHD | Sin búsqueda externa reaparece una solución MHD espectacular: campos >10T, presión oceánica y deflexión Lorentz. Es creativa, pero muy vulnerable a mezcla turbulenta y escala física. | Ejecución 0098, id 1. Desalinización, semilla 43, repetición 08. F=2, Q=0,806. |
| Apéndice | no_agent4 | Polarización microfluídica de concentración iónica | La ausencia del último agente no impide producir una hipótesis sólida: reduce MHD macroscópico a electrocinética microfluídica, una ruta físicamente más defendible. | Ejecución 0022, id 3. Desalinización, semilla 42, repetición 02. F=3, Q=0,917. |
| Apéndice | no_lateral_lenses | Esponja electroquímica de iones basada en análogos de azul de Prusia | Al eliminar lentes laterales aparece una solución menos exótica y más cercana a tecnologías existentes de intercalación electroquímica. Gana plausibilidad, pero pierde parte del salto combinatorio. | Ejecución 0042, id 1. Desalinización, semilla 42, repetición 02. F=3, Q=0,917. |
| Apéndice | no_search | Mercado cuadrático de tiempo para controlar la agenda | Propone una economía procedimental donde el coste de acaparar tiempo parlamentario crece cuadráticamente. Es elegante, aunque depende de una formalización institucional no contrastada externamente. | Ejecución 0155, id 1. Parlamento, semilla 42, repetición 05. F=3, Q=0,917. |
| Apéndice | no_agent4 | Activador del artículo 85.1 con conocimiento cero | Ejemplo especialmente valioso: usa un anclaje jurídico local, el umbral de 70 firmas del artículo 85.1, y aplica criptografía solo para proteger coordinación interna. | Ejecución 0150, id 1. Parlamento, semilla 42, repetición 10. F=3, Q=0,917. |
| Apéndice | no_lateral_lenses | Dividendo procedimental ponderado por consenso | Solución sobria y probablemente implementable: recompensa el consenso multipartidista con vía rápida procedimental. Muestra que retirar lateralidad puede aumentar claridad institucional. | Ejecución 0167, id 2. Parlamento, semilla 42, repetición 07. F=3, Q=0,917. |
为什么重要
这提示我们,与其一味压制大语言模型的幻觉,不如用批评、过滤和评估把它约束起来,变成产生研究点子的工具。不过这只是一项规模有限、探索性的研究——仅有三个问题-种子组合、180次运行,由资源有限的高中生独立完成,作者自己也强调结论只是方向性的,不是严格的统计证明。
本文术语
- 幻觉(hallucination) · 语言模型煞有介事地编造出不真实或没有依据的内容
- 多智能体架构 · 由多个分工不同的AI模块相互协作或制衡的系统,而非单一模型独自完成任务
- 认识论摩擦(Epistemological Friction) · 本文提出的概念,指自由发散的生成智能体与核实批评的评估智能体之间的对抗过程
- 自我反思(self-reflection) · 让模型对自己先前的输出进行检查和修改的方法
- 结构有效性(structural validity) · 本文用来判断一个假设是否具备明确机制、可测量变量、可检验性且不与其他假设重复的综合标准
论文原文摘要(英文)
Contemporary Large Language Models (LLMs) are increasingly aligned to suppress hallucinations, prioritizing factual retrieval over combinatorial creativity. While crucial for mitigating misinformation, this alignment may also restrict speculative Research and Development (R&D) by encouraging what this work operationally treats as semantic overfitting and diversity collapse. In this paper, we propose a Rust-based multi-agent orchestration that uses the contrast between narrative daydreaming and executive control as a functional analogy, not as a neurocognitive claim. The system instigates an Epistemological Friction loop between a high-entropy generating agent and a web-grounded evaluating agent, mediated by a low-entropy semantic bottleneck intended to reduce noise and repetition. Initial experiments generated diverse, viability-rated hypotheses across physical and social-science domains. We additionally report an exploratory paired baseline and ablation study comparing the full system against direct prompting, self-reflection, removal of the semantic filter, removal of search grounding, and removal of lateral lenses. The results place direct prompting among the weakest conditions across most observed metrics, but they do not show a general superiority of the full system over simple self-reflection. Instead, they suggest that each architecture shifts the balance between originality, feasibility, diversity, and empirical grounding in different ways, and that the full system provides its main advantages when hypotheses must survive strong physical, empirical, or institutional constraints. These findings do not show that hallucination is useful in isolation; they suggest that speculative generation gains value only when constrained by architecture, empirical grounding, and explicit evaluation.
在 arXiv 阅读最新论文
- LoRA-GA$^2$: Low Rank Adaptation with Multi-step Gradient Adaptive Alignment在正式微调前先偷看几步训练的梯度,让LoRA的初始化更聪明
- Remember, Verify, or Ask? Cross-Family Evaluation of Memory Commitment in LLM AgentsAI助手在该向你提问的时候,却更愿意自己去核实事实
- Robust Incomplete Multimodal Sentiment Analysis via Iterative Proxy Correction文本信息缺失或损坏时,这个AI不靠一次性猜测,而是反复修正猜测结果,从而更准确地判断情绪
- Generating Diverse Personas for User Simulators to Test Interview Dialogue Systems要测试访谈式对话系统需要大量不同性格的虚拟用户,这项研究用大语言模型自动生成这些虚拟用户人设
- Rethinking Patch Based Multivariate Time Series Forecasting with Semantic Structured Partitioning别再机械切分时间序列,按语义把它切成有意义的块
- Reliable Financial Named Entity Recognition under Domain ShiftAI在正式文件里学到的自信,一到推特上就变得不可信
- Bringing analytic rigor to agentic AI for science: The Brain Researcher platform for neuroimaging data analysis让AI分析脑影像数据时,把“为什么这个结论可信”也一并记录下来
- GenMatch: An End-to-End Generative Matching Framework for Micro-View Order-Dispatching in Ride-Hailing滴滴把打车派单从预测-计算-匹配三段式流程改成一次生成完成,线上效果提升明显
METAL LAB 最新报道
图片来源: Nicolas Rodriguez-Alvarez (IES Parquesol et al., arXiv:2608.19206, CC BY 4.0