{"path":"inductive-reasoning-research.md","content":"# Inductive Reasoning in Argumentation: State of the Art (2023-2026)\n\nResearch survey for Deliberus platform design. Covers advances in LLM probabilistic reasoning, Bayesian argumentation, argument strength quantification, and practical implementations.\n\n**The core problem**: Most real-world arguments are inductive — based on evidence, statistics, probability, analogy, generalization. Deductive arguments are structurally checkable; inductive arguments require assessing *degrees* of support. How do recent advances make this tractable for a deliberation platform?\n\n---\n\n## 1. LLMs as Bayesian Reasoners\n\n### Can LLMs do calibrated probabilistic reasoning?\n\n**Short answer**: Partially, and improving fast. LLMs approximate Bayesian updating under the right conditions, but have systematic biases.\n\n**Key papers**:\n\n- **\"Bayesian Teaching Enables Probabilistic Reasoning in Large Language Models\"** (Qiu et al., Nature Communications, Jan 2026). Google Research showed that LLMs trained to mimic optimal Bayesian models can construct world representations and form probabilistic beliefs. The method (\"Bayesian teaching\") trains LLMs to predict like an optimal Bayesian agent. Published in Nature Communications — the strongest endorsement yet that LLMs can be made to reason probabilistically. [arxiv.org/html/2503.17523v3](https://arxiv.org/html/2503.17523v3)\n\n- **\"Enough Coin Flips Can Make LLMs Act Bayesian\"** (Gupta et al., ACL 2025). UC Berkeley found that with sufficient in-context examples, LLMs update priors in a Bayesian manner. Key insight: deviations from Bayesian behavior come primarily from miscalibrated *priors*, not from flawed *update mechanisms*. This matters for Deliberus: if we supply good priors (community base rates, established evidence), LLMs can do reasonable posterior updates. [aclanthology.org/2025.acl-long.377](https://aclanthology.org/2025.acl-long.377/)\n\n- **\"From Evidence to Belief: A Bayesian Epistemology Approach to Language Models\"** (Kim et al., NAACL 2025). Found that LLMs follow the Bayesian confirmation assumption well with *true* evidence but fail with other evidence types. LLMs are biased toward \"golden evidence\" (clean, annotated data) and handle the gradient of relevance/irrelevance poorly. [arxiv.org/abs/2504.19622](https://arxiv.org/abs/2504.19622)\n\n- **\"QUITE: Quantifying Uncertainty in Natural Language Text in Bayesian Reasoning Scenarios\"** (Schrader et al., EMNLP 2024). Created a benchmark of real-world Bayesian reasoning with natural language premises. Finding: logic-based (neuro-symbolic) models outperform pure LLMs on causal, evidential, and explaining-away reasoning. [aclanthology.org/2024.emnlp-main.153](https://aclanthology.org/2024.emnlp-main.153/)\n\n### LLM Forecasting Accuracy\n\n- **ForecastBench** (Karger et al., 2024-2026). Dynamic benchmark comparing LLM forecasting to superforecasters. Current state: best LLM (GPT-4.5) achieves Brier score 0.101 vs superforecasters' 0.081. LLMs already outperform non-expert crowds. Projected LLM-superforecaster parity: **November 2026** (95% CI: Dec 2025 - Jan 2028). [forecastbench.org](https://www.forecastbench.org/)\n\n- **FutureEval** (Metaculus, Feb 2026). Metaculus launched a dedicated benchmark for AI forecasting accuracy, institutionalizing the comparison between AI and human forecasters. [metaculus.com/notebooks/42225](https://www.metaculus.com/notebooks/42225/futureeval-measures-the-forecasting-accuracy-of-ai/)\n\n- **EvolveCast** (Yuan et al., 2026). Framework for evaluating whether LLMs *update* their forecasts appropriately with new information — testing Bayesian updating behavior specifically. [arxiv.org/html/2509.23936v2](https://arxiv.org/html/2509.23936v2)\n\n- **Mantic / Thinking Machines Lab** (Mar 2026). Training LLMs specifically for prediction, approaching superforecaster-level accuracy on geopolitics and current affairs. [thinkingmachines.ai/news/training-llms-to-predict-world-events](https://thinkingmachines.ai/news/training-llms-to-predict-world-events/)\n\n**Implication for Deliberus**: LLMs are credible probability estimators for well-defined factual questions. For the \"how likely is this claim true?\" component, an LLM-based estimator calibrated against prediction market track records could provide initial probability anchors that communities then adjust.\n\n---\n\n## 2. Bayesian Argumentation Frameworks\n\n### Extending Dung with Probabilities\n\nThe formal argumentation community has been actively extending abstract argumentation with probabilistic reasoning:\n\n- **Probabilistic Argumentation Frameworks (PrAFs)**: Bistarelli, David, Santini & Taticchi (2024-2025) published multiple papers extending Dung's frameworks with probabilities, including temporal duration-based variants and fast computation algorithms for acyclic graphs. [inria.hal.science/hal-04851343](https://inria.hal.science/hal-04851343v1/document)\n\n- **\"From Networks to Narratives: Bayes Nets and the Problems of Argumentation\"** (Keshmirian et al., RATIO 2024). Bridges Bayesian networks and argumentation directly — analyzes what Bayes nets can and cannot capture about argumentative reasoning. [link.springer.com/chapter/10.1007/978-3-031-63536-6_14](https://link.springer.com/chapter/10.1007/978-3-031-63536-6_14)\n\n- **\"Argumentation-Based Probabilistic Causal Reasoning\"** (Bengel et al., RATIO 2024). Integrates causal reasoning into argumentation frameworks using probabilistic methods. [link.springer.com/chapter/10.1007/978-3-031-63536-6_13](https://link.springer.com/chapter/10.1007/978-3-031-63536-6_13)\n\n- **\"Probabilistic Causal Bipolar Abstract Argumentation\"** (Morveli-Espinoza et al., 2023). Uses credal networks to handle uncertainty in bipolar argumentation (where arguments can both attack AND support each other). This bipolar model is more realistic for real-world deliberation where evidence can partially support and partially undermine. [link.springer.com/content/pdf/10.1007/s10472-023-09851-4.pdf](https://link.springer.com/content/pdf/10.1007/s10472-023-09851-4.pdf)\n\n### Quantitative Bipolar Argumentation Frameworks (QBAFs)\n\nQBAFs are emerging as the dominant formal structure for probabilistic argumentation. Arguments have numeric strength, and attack/support relations modify that strength via \"gradual semantics.\" Multiple 2025-2026 systems build on QBAFs (see ArgRAG, ARGORA, MArgE below).\n\n**Implication for Deliberus**: QBAFs are the most promising formal backbone. They naturally handle: (1) arguments with varying strength, (2) both attack and support relations, (3) deterministic computation of final argument strength from the graph structure. A Deliberus \"argument kernel\" could be a QBAF with user-contributed strength estimates.\n\n---\n\n## 3. Argument Strength Quantification & Evidence Weighing\n\n### BIRD: Bayesian Inference for Reliable Decisions\n\n**\"BIRD: A Trustworthy Bayesian Inference Framework for Large Language Models\"** (Feng et al., ICLR 2025 Oral). Decomposes LLM decision-making into three phases: (1) **Abduction** — LLM identifies relevant factors for a claim, (2) **Entailment** — LLM checks which factors are supported by evidence, (3) **Deduction** — external Bayesian model computes calibrated probabilities. Outperforms GPT-4 by 35% on probability calibration using open-source Llama models. [arxiv.org/abs/2404.12494](https://arxiv.org/abs/2404.12494)\n\nThis three-phase decomposition (abduction → entailment → deduction) is directly applicable to Deliberus: users contribute the abductive factors, the platform checks entailment via NLI/LLM, and a Bayesian model computes calibrated strength.\n\n### Probabilistic Soundness in Reasoning Chains\n\n**\"Probabilistic Soundness Guarantees in LLM Reasoning Chains\"** (You et al., EMNLP 2025). Introduces ARES (Autoregressive Reasoning Entailment Stability), which evaluates each reasoning step based on previously-verified premises. Achieves 90.3% F1 on detecting propagated errors in long reasoning chains. [aclanthology.org/2025.emnlp-main.382](https://aclanthology.org/2025.emnlp-main.382/)\n\n**Implication for Deliberus**: When users construct multi-step inductive arguments, the platform can validate each step's entailment from its premises and flag where reasoning chains become unreliable. This turns \"this argument feels weak\" into \"step 3 has only 0.42 entailment probability from its premises.\"\n\n### BayesAgent: Verbalized Probabilistic Graphical Models\n\n**\"BayesAgent\"** (Shen et al., AAAI 2026). LLMs construct probabilistic graphical models (PGMs) in natural language — identifying latent variables and their dependencies, then performing Bayesian inference. No domain expertise required for model construction. [arxiv.org/abs/2406.05516](https://arxiv.org/abs/2406.05516)\n\n**Implication for Deliberus**: An LLM could automatically construct a probabilistic causal model from a user's argument, making the implicit reasoning structure explicit and computable.\n\n---\n\n## 4. Decomposing Inductive Arguments\n\n### Atomic Claim Extraction\n\n**Claimify** (Microsoft Research, Feb 2025). Three-stage pipeline: Selection → Disambiguation → Decomposition. Extracts atomic, verifiable, decontextualized claims from text with 99% entailment accuracy and 87.6% coverage. Critical feature: it identifies when text has multiple possible interpretations and only extracts claims when confidence is high. [microsoft.com/en-us/research/blog/claimify](https://www.microsoft.com/en-us/research/blog/claimify-extracting-high-quality-claims-from-language-model-outputs/) | Paper: [arxiv.org/abs/2502.10855](https://arxiv.org/abs/2502.10855)\n\nUnofficial open implementation: [github.com/deshwalmahesh/claimify](https://github.com/deshwalmahesh/claimify) — LLM-agnostic, accepts any `llm(prompt, temperature) -> str` function.\n\n### Question Decomposition\n\n**Iterated Decomposition** (Ought/Elicit, Reppert et al., 2023). Foundational work on breaking complex research questions into assessable sub-questions. The key insight: decomposition improves faithfulness of model-generated reasoning because it forces explicit intermediate steps. [arxiv.org/pdf/2301.01751v2](https://export.arxiv.org/pdf/2301.01751v2.pdf)\n\n**Anthropic's Question Decomposition** (Radhakrishnan et al., 2023). Demonstrated that decomposing questions improves faithfulness of model-generated reasoning. [anthropic.com research paper](https://www-cdn.anthropic.com/8154fb1d828cdc390dc1fa442d84034948679c47/question-decomposition-improves-the-faithfulness-of-model-generated-reasoning.pdf)\n\n### Analysis of Competing Hypotheses (ACH)\n\nACH remains relevant as a structured analytic technique. A Bayesian probabilistic framework for ACH evaluation shows how ACH tables can be represented as Bayesian networks, with the BN providing additional generality for handling dependent evidence. [cse.sc.edu/~mgv/reports/IA-05.pdf](https://cse.sc.edu/~mgv/reports/IA-05.pdf)\n\n**Implication for Deliberus**: The decomposition pipeline for a Deliberus claim could be: (1) Claimify extracts atomic claims from user arguments, (2) each claim gets an independent probability estimate, (3) claim dependencies form a Bayesian network or QBAF, (4) the overall argument strength is computed from the composed structure. This makes \"climate change is caused by human activity\" tractable by decomposing into sub-claims like \"CO2 levels have risen 50% since pre-industrial times\" (highly certain), \"CO2 is a greenhouse gas\" (established physics), \"temperature rise correlates with CO2\" (strong evidence), etc.\n\n---\n\n## 5. Scientific Method as Algorithm\n\n### POPPER: Automated Hypothesis Validation\n\n**POPPER** (Huang, Jin, Li et al., ICML 2025). Named after Karl Popper, this Stanford framework automates hypothesis validation through agentic sequential falsification. Key technical innovation: uses sequential e-values with strict Type-I error control. An LLM agent designs falsification experiments, a ReAct agent executes them, p-values are aggregated into sequential e-values. Compared to human scientists, achieves comparable performance while reducing time 10x. Open source. [github.com/snap-stanford/POPPER](https://github.com/snap-stanford/POPPER) | [arxiv.org/abs/2502.09858](https://arxiv.org/abs/2502.09858)\n\n### AutoDiscovery: Bayesian Surprise for Scientific Discovery\n\n**AutoDiscovery** (Allen AI, NeurIPS 2025). Uses Bayesian surprise (measuring how much data shifts posterior from prior) to guide open-ended scientific discovery. Monte Carlo Tree Search navigates the hypothesis space. Already deployed in AstaLabs for cancer genomics, marine ecology, and social science. [allenai.org/blog/autodiscovery](https://allenai.org/blog/autodiscovery) | [arxiv.org/abs/2507.00310](https://arxiv.org/abs/2507.00310)\n\n### Evidence and Elimination: Bayesian Falsification\n\n**\"Evidence and Elimination: A Bayesian Interpretation of Falsification in Scientific Practice\"** (Costa, Dec 2025). Formally reconciles Popper's falsification with Bayesian updating — showing that falsification is a special case of Bayesian evidence accumulation where evidence drives posterior probability toward zero. [arxiv.org/abs/2512.06777](https://arxiv.org/abs/2512.06777)\n\n**Implication for Deliberus**: The Popper-Bayes reconciliation provides the theoretical foundation for Deliberus's approach to inductive arguments: claims aren't \"proven\" but accumulated evidence drives their posterior probability up or down. The platform could track the \"falsification history\" of each claim — what evidence has been brought against it and how it survived. POPPER's sequential e-value framework could provide formal statistical guarantees for when a claim has been \"sufficiently tested.\"\n\n---\n\n## 6. Argumentation + LLMs: The New Synthesis (2024-2026)\n\nThis is the most active and directly relevant research area. Multiple groups are independently converging on the same architecture: **LLM argument extraction + structured argumentation frameworks + deterministic reasoning**.\n\n### ARGORA: Orchestrated Argumentation with Causal Grounding\n\n(Jin et al., Jan 2026). Multi-expert LLM discussions organized into explicit argumentation graphs. Key innovation: interprets QBAFs as **structural causal models**, enabling counterfactual analysis — \"would the conclusion change if we removed argument X?\" This provides genuine explainability. [arxiv.org/abs/2601.21533](https://arxiv.org/abs/2601.21533)\n\n### ArgRAG: Argumentation-Enhanced RAG\n\n(Zhu et al., NeSy 2025). Replaces black-box RAG reasoning with structured QBAFs. Retrieved evidence becomes arguments with support/attack relations, evaluated by gradual semantics. Achieves highest accuracy across all LLM backbones on fact verification while being fully transparent. [proceedings.mlr.press/v284/zhu25a](https://proceedings.mlr.press/v284/zhu25a.html)\n\n### MArgE: Multi-LLM Argumentative Evidence Meshing\n\n(Ng, Jiang, Freedman, Rago, Toni, Aug 2025). Multiple LLMs independently generate structured pro/con argument trees for a claim. Trees are merged into a unified structure, scored for intrinsic strength, evaluated via gradual semantics. Significantly outperforms single LLMs and unstructured debate. [arxiv.org/abs/2508.02584](https://arxiv.org/abs/2508.02584)\n\n### SAFE: Structured Argumentation for Fact-checking with Explanations\n\n(Wang, Cabrio, Villata, IJCAI 2025). Uses computational argumentation explicitly for fact-checking, providing structured explanations. [ijcai.org/proceedings/2025/1274.pdf](https://www.ijcai.org/proceedings/2025/1274.pdf)\n\n### LLM Argument Mining + Description Logics\n\n(Mar 2026). Unified framework: LLM-based argument mining → fuzzy argumentative knowledge base → quantitative argumentation semantics → fuzzy description logic for query answering. First to combine argument mining, formal argumentation, and ontological reasoning. [arxiv.org/abs/2603.02858](https://arxiv.org/abs/2603.02858)\n\n### Retrieval- and Argumentation-Enhanced Multi-Agent LLMs for Forecasting\n\n(Oct 2025). Multi-agent framework with ArgLLM, Relation-based Argument Mining, and RAG-ArgLLM agents, each contributing evidence as QBAFs. Three-agent configurations improve forecasting accuracy. [arxiv.org/abs/2510.24303](https://arxiv.org/abs/2510.24303)\n\n### Multi-Agent Debate for Claim Verification\n\n- **Tool-MAD** (Jeong et al., Jan 2026). Multi-agent debate with diverse tool augmentation for fact verification. [arxiv.org/pdf/2601.04742](https://arxiv.org/pdf/2601.04742)\n- **MAD-Fact** (Lin et al., Oct 2025). Multi-agent debate for long-form factuality evaluation. [arxiv.org/html/2510.22967v1](https://arxiv.org/html/2510.22967v1)\n- **Argus** (open source). Multi-agent AI framework where LLMs debate to verify claims using Bayesian reasoning, adversarial evidence, and calibrated verdicts with full audit trails. [github.com/Argus-Framework/argus-ai-debate](https://github.com/Argus-Framework/argus-ai-debate)\n\n**Implication for Deliberus**: The converging architecture is clear — LLMs mine arguments from evidence and user contributions, structure them as QBAFs, and deterministic gradual semantics compute final strength. Deliberus can adopt this pattern where: (1) users contribute arguments and evidence, (2) LLMs extract atomic claims and identify attack/support relations, (3) a QBAF engine computes argument strength, (4) ARGORA-style causal analysis explains *why* conclusions hold. The multi-LLM debate pattern (MArgE) could power a \"devil's advocate\" feature that automatically generates counterarguments.\n\n---\n\n## 7. Epistemic Calibration & Uncertainty Presentation\n\n### Superforecasting Methods\n\nTetlock's superforecasting methodology remains the gold standard for calibrated probability estimation:\n- Break questions into sub-components\n- Start from base rates (outside view)\n- Update incrementally with specific evidence (inside view)\n- Track calibration over time\n\n**Good Judgment Open** and **Metaculus** operationalize these principles at scale. [goodjudgment.com/services/online-training](https://goodjudgment.com/services/online-training/)\n\n### Calibration Training\n\n**Open Philanthropy Calibration App** (2018, now unmaintained). Demonstrated that calibration training improves probability estimation. The key finding: most people are overconfident and calibration training corrects this. [openphilanthropy.org/research/new-web-app-for-calibration-training](https://www.openphilanthropy.org/research/new-web-app-for-calibration-training/)\n\n### Uncertainty Visualization Research (2024-2025)\n\nIEEE Workshops on Uncertainty Visualization (2024, 2025) identified key findings:\n- **Density strips** produce the best alignment between user perception and actual probabilities, with shortest response times and highest cognitive arousal\n- Visualizations engage perceptual processes and are less likely to overload cognition than numbers or words\n- Non-experts find it challenging to perceive reliability of probabilistic systems accurately\n- The gap between perceived and actual probability remains an active research problem\n\n[tusharathawale.github.io/uncertainty-vis-workshop-2025](https://tusharathawale.github.io/uncertainty-vis-workshop-2025/) | [onlinelibrary.wiley.com/doi/full/10.1002/for.3222](https://onlinelibrary.wiley.com/doi/full/10.1002/for.3222)\n\n**Intelligence analysis research** (2025) found that visualizing uncertainty (vs verbalizing) significantly affects analyst confidence and decision quality. [tandfonline.com/doi/full/10.1080/02684527.2025.2468049](https://www.tandfonline.com/doi/full/10.1080/02684527.2025.2468049)\n\n**Implication for Deliberus**: The UI should use density strips or similar visual encodings for probability, not just numbers. Calibration training could be built into the onboarding flow — users who estimate probabilities more accurately contribute more value. The platform could track each user's calibration score (like Metaculus does) and weight their probability estimates accordingly.\n\n---\n\n## 8. AI-Assisted Deliberation Platforms\n\n### The Habermas Machine\n\n(Tessler et al., Science, Oct 2024). AI mediator for group deliberation that outperforms human mediators at generating consensus statements (56% vs 44% preference). Key finding: AI mediation genuinely reduces intra-group divisions rather than just finding lowest-common-denominator statements. [science.org/doi/10.1126/science.adq2852](https://www.science.org/doi/10.1126/science.adq2852)\n\n### Elicit's Vision\n\nAndreas Stuhlmuller (Elicit CEO, Oct 2025): \"Can we build systems that help humans reason better — and can we do it fast enough to matter?\" Elicit's approach: decompose research questions, supervise reasoning processes, make reasoning transparent. The first AI-written Community Note on X was accepted — scaling this to every post across social media is the goal. [elicit.com/blog/ai-for-human-reasoning](https://elicit.com/blog/ai-for-human-reasoning)\n\n### Kialo\n\nThe world's largest argument mapping platform. Hierarchical pro/con debate trees, crowdsourced argument quality ratings. Primarily used in education (HundrED Global Collection 2025, AASL Best Digital Tools 2025). Limitations: no probabilistic reasoning, no formal argument strength computation, purely structural. [kialo-edu.com](https://www.kialo-edu.com/)\n\n### Interoperable Deliberative Tools\n\nMetagov is developing an RFP for interoperable deliberative tools — standardized formats for exchanging argumentation structures between platforms. [metagov.github.io/interop](https://metagov.github.io/interop/)\n\n### AI-Enhanced Deliberative Democracy\n\n(Revel & Penigaud, 2025). Comprehensive analysis of AI methods for democratic deliberation: opinion mining, argument extraction, consensus detection, AI-powered summarization. [arxiv.org/pdf/2503.05830](https://arxiv.org/pdf/2503.05830)\n\n### Human-AI Deliberation\n\n(CHI 2025). Design and evaluation of LLM-empowered deliberative AI for decision-making. Found that AI assistance improves decision quality when it makes its reasoning transparent. [dl.acm.org/doi/10.1145/3706598.3713423](https://dl.acm.org/doi/10.1145/3706598.3713423)\n\n---\n\n## 9. Synthesis: What This Means for Deliberus\n\n### The Convergent Architecture\n\nMultiple independent research threads are converging on the same pattern:\n\n```\nUser Arguments → LLM Claim Extraction → Structured Argumentation Graph →\nBayesian/Probabilistic Strength Computation → Causal Explanation\n```\n\nThe components exist. What doesn't exist is a platform that puts them together for *collaborative public deliberation* rather than automated fact-checking.\n\n### Proposed Technical Stack for Inductive Arguments\n\n1. **Claim Decomposition Layer**: Claimify-style atomic claim extraction from user arguments. Each claim tagged as factual (verifiable), normative (value judgment), or definitional.\n\n2. **Evidence Attachment**: Users and AI agents attach evidence to claims. Evidence quality scored by LLM (source reliability, statistical rigor, sample size, potential confounders).\n\n3. **Argument Graph Engine**: QBAFs with gradual semantics. Support and attack relations between claims. Deterministic computation of argument strength from graph structure.\n\n4. **Bayesian Probability Layer**: Each factual claim has a community probability estimate (aggregated from user estimates, weighted by calibration score). Bayesian updating as new evidence arrives. BIRD-style decomposition: abduction → entailment → deduction.\n\n5. **Falsification Tracking**: POPPER-inspired sequential testing. Claims that survive more falsification attempts earn higher credibility. \"Battle-tested\" badge for claims with high e-values.\n\n6. **Causal Explanation Engine**: ARGORA-style counterfactual analysis. \"This conclusion depends critically on claim X — if X were removed, the overall argument strength drops from 0.82 to 0.31.\"\n\n7. **Calibration System**: User calibration scores tracked over time (Metaculus-style). Better-calibrated users' estimates weighted more heavily. Calibration training built into onboarding.\n\n8. **Uncertainty Visualization**: Density strips for probabilities. Color-coded argument strength. Visual indicators of where an argument is weakest.\n\n### What's Now Possible That Wasn't in 2012\n\n| 2012 Problem | 2026 Solution |\n|---|---|\n| \"Can't decompose arguments automatically\" | Claimify, LLM argument mining |\n| \"Can't estimate claim probabilities\" | LLMs approaching superforecaster calibration |\n| \"Formal argumentation is too abstract\" | QBAFs with gradual semantics, natural language grounding |\n| \"Can't explain why an argument is strong/weak\" | ARGORA causal counterfactuals |\n| \"No way to combine probability with argumentation\" | Probabilistic AFs, BIRD, BayesAgent |\n| \"Can't generate counterarguments automatically\" | Multi-agent debate (MArgE, MAD-Fact) |\n| \"Scientific method can't be automated\" | POPPER, AutoDiscovery |\n| \"People are bad at probability\" | Calibration training, Metaculus track record |\n\n### Open Research Questions for Deliberus\n\n1. **Normative claims**: All the above handles *factual* claims well. How to handle \"we *should* do X\"? This requires a value/fact boundary that no current system handles cleanly.\n\n2. **Composing probabilities across heterogeneous evidence types**: A statistical study, an expert opinion, and a historical analogy all support a claim — how to combine them formally?\n\n3. **Adversarial robustness**: If users game probability estimates, how does the platform detect and correct this? Prediction market mechanisms (betting with reputation) are one approach.\n\n4. **Cognitive load**: The full formal machinery could overwhelm users. How much can be hidden behind a \"simple mode\" UI while maintaining intellectual honesty?\n\n5. **The prior problem**: LLMs have miscalibrated priors (Berkeley coin flip paper). What serves as the \"prior\" for a brand-new claim on Deliberus? Base rates from prediction markets? Expert estimates? Community polls?\n\n---\n\n## Key References (Chronological)\n\n| Year | Paper/System | Venue | Relevance |\n|------|---|---|---|\n| 2023 | Iterated Decomposition (Ought/Elicit) | arXiv | Question decomposition methodology |\n| 2023 | Probabilistic Causal Bipolar Argumentation | Ann. Math AI | Credal networks + argumentation |\n| 2024 | QUITE | EMNLP | Bayesian reasoning benchmark |\n| 2024 | BIRD | ICLR 2025 Oral | Bayesian inference framework for LLMs |\n| 2024 | Bayes Nets and Argumentation (RATIO) | Springer | Bridging Bayes nets and argumentation |\n| 2024 | Habermas Machine | Science | AI-mediated group deliberation |\n| 2024 | ForecastBench | arXiv | LLM vs superforecaster benchmarks |\n| 2025 | Enough Coin Flips | ACL | LLMs as Bayesian updaters |\n| 2025 | From Evidence to Belief | NAACL | Bayesian epistemology for LMs |\n| 2025 | POPPER | ICML | Automated hypothesis falsification |\n| 2025 | ArgRAG | NeSy | Argumentation-enhanced RAG |\n| 2025 | MArgE | arXiv | Multi-LLM argumentative evidence |\n| 2025 | ARES | EMNLP | Probabilistic soundness in reasoning |\n| 2025 | AutoDiscovery | NeurIPS | Bayesian surprise for discovery |\n| 2025 | Claimify | Microsoft Research | Atomic claim extraction |\n| 2025 | SAFE (argumentation) | IJCAI | Structured fact-checking |\n| 2025 | Evidence and Elimination | arXiv | Popper-Bayes reconciliation |\n| 2026 | Bayesian Teaching | Nature Comms | Training LLMs as Bayesian agents |\n| 2026 | ARGORA | arXiv | Causal argumentation frameworks |\n| 2026 | BayesAgent | AAAI | Verbalized probabilistic graphical models |\n| 2026 | FutureEval | Metaculus | AI forecasting benchmark |\n| 2026 | LLM Arg Mining + DL | arXiv | Unified argumentation + ontology |\n| 2026 | Tool-MAD | arXiv | Multi-agent debate for verification |\n| 2026 | Argus | GitHub | Open-source claim verification via debate |\n"}