Artificial Intelligence · arXiv:2609.39341 v1 · submitted 30 Sep 2026 · 13 references
Understanding as No-Arbitrage: Bounded Dutch Books as a Definition and Training Objective for Language Models
The paper says
“Does a language model merely predict tokens, or does it understand what it says? We make this question measurable by defining "understanding" through the lens of no-arbitrage.”
more of the abstract
A model understands a vocabulary to a certain degree if a computationally bounded trader cannot extract guaranteed profit by betting against the model's probabilities on logically related claims (a "Dutch book"). We establish three theoretical results: first, because full logical coherence is computationally intractable, understanding is inherently graded, not absolute. Second, we prove that the exact optimum of standard next-token prediction is inherently incoherent across different question formats; the flaw lies in the training objective, not the architecture. Third, we show that uncertainty accumulates predictably along reasoning chains, making unjustified overconfidence an arbitrage opportunity in itself. To address this, we introduce Arbitr, a training framework where an adversarial trader penalizes the model for logical inconsistencies, paired with a calibration anchor to prevent uninformative collapse. Across five pre-registered experiments on Qwen2.5 and Phi-3.5 models, we demonstrate that standard models are highly exploitable across different phrasings. Arbitr reduces this exploitability by orders of magnitude without sacrificing task accuracy, and the effect successfully transfers to unseen logical patterns and new model families. Crucially, we uncover a scaling illusion: at 7B parameters, near-zero measured incoherence often coincides with extreme, unjustified confidence. We conclude that while Arbitr enforces rigorous logical consistency, coherence is a necessary condition for knowledge, but not a sufficient one
from the abstract
Pith says
Machine review · deepseek-v4-flash · 1 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Solid, honestly reported theory-plus-training paper whose load-bearing soft spot is that every measured reduction runs through the same fragile first-token readout, so the verdict is conditional until a third-format or independent readout settles it. This line is wrong
3 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 13 references say.
The paper says, at §6–§7, Tables 1–4
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.84):
“Elicitation is fragile: credences come from the first answer-token, and three silent readout failures happened during the campaign, all caught by cross-format books rather than the coverage filter.”
Pith says: Objection 1 of 3 · major
Every Inc measurement and every Arbitr gradient is computed from the first-token answer distribution, and the paper itself documents three silent readout failures during the campaign, including one that inverted answers at high coverage, plus a numbered-choice format only weakly correlated with the other two ([b233]).
more of objection 1
Because the cross-format books that caught those failures are the same law-of-one-price quantity used both as the main measurement axis and as a training target, the readout is not independently validated. A held-out third elicitation format, a sampled-answer readout, or a semantic-clustering readout of the kind the authors already suggest in [b234] would be needed to rule out the possibility that the headline reductions in Tables 3–4 are readout-level effects rather than changes in the model's underlying credences.
The paper says, at §7.1–7.3, Table 4
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.82):
“Wenow reportthe negation reduction decomposed rather thanasa single multiplier, 12 and read this as a partial, disclosed confirmation of the worry that the attribution to the trader was unestablished: real on law-of-one-price, overstated if left unstated on negation.”
Pith says: Objection 2 of 3 · major
The primary registered endpoints P1 and P2 are reductions in quantities that the Arbitr objective explicitly optimizes: the negation book |p(phi)+p(not-phi)-1|/2 and the cross-format law-of-one-price book |p_F1(phi)-p_F2(phi)|/2, as described in §5 and §7. Part of the negation reduction is enforced by construction, as the E-A3 ablation concedes: the calibration anchor alone, which supervises p(not-phi)=1-t, accounts for a factor of about 2.4 ([b198]).
more of objection 2
For the LOP channel the trader appears to do the work, but those held-out books are unsupervised and are scored from the same first-token readout, so readout-local format-invariance training remains a live alternative. The paper should provide readout-independence evidence, or it should restrict the claim to readout-local consistency rather than belief-level coherence.
The paper says, at §7.3
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.78):
“Unseen-template discrimination.The original E-A1b adapters (arms A and B0.3) no longer exist as saved weights—only their aggregate metrics survived the move to this machine—so the discrimination test could not be run against the trained trader arm directly; this is logged as a resource gap, not glossed over.”
Pith says: Objection 3 of 3 · major
The direct unseen-template discrimination test on the trained trader arm could not be run on the Qwen2.5 models whose headline results are reported because the adapters no longer exist ([b208]); it was run instead on the Phi-3.5 replication, and the Qwen-side check used only the anchor-only AN arm.
more of objection 3
The abstract's unqualified claim that the effect 'transfers to unseen logical patterns and new model families' therefore rests on the cross-family replication plus an untrained level-3 transitivity pattern, not on the original model family. This should be stated where the claim is made, and a re-trained Qwen replica should be added before the claim is made without qualification.
4 minor objections
The paper says, at §7.3
Pith says: minor objection 1 of 4
The text refers to 'Table eftab:ea0', an unresolved LaTeX label; it should read Table 1.
The paper says, at §7.3
Pith says: minor objection 2 of 4
The abbreviation C2 is used ('against the surface-trick reading of C2') but never defined; either define the criterion or remove the reference.
show 2 more
The paper says, at §5
Pith says: minor objection 3 of 4
The Arbitr objective is described verbally; a displayed equation for the exact trader's profit, the anchor Brier term, and the checkpoint composite would improve reproducibility.
The paper says, at §6, Table 1
Pith says: minor objection 4 of 4
The cross-format mean profits in the bottom half of Table 1 are reported without confidence intervals, although the primary row has them; adding intervals would make the scale comparison easier to assess.
What would settle it
Re-run the E-A1b evaluation (arms A control and B0.3) on the held-out real-content books (206 books, 412 propositions) but elicit credences with an independent readout never used in training: semantic-cluster averaging over sampled open-form answers following Farquhar et al. 2024 - the authors themselves name this 'the right foundation' ([b234]) - or a third linguistic format (e.g., English yes/no or French) absent from Arbitr's training batches. If negation Inc and cross-format LOP profit measured through the independent readout on the B0.3 arm stay near the first-token values (about 0.0015 and 0.041) and below control, the readout-artifact worry is defused and the reduction reflects genuine belief coherence. If they revert toward control's 0.28 and 0.12, the headline reduction is a first-token artifact.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 1 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Two things before you spend the hour. The theoretical core is real: the objective-defect theorem (Thm 7) is a genuine closed-form result, on a mixed-author corpus the exact NLL optimum prices phi at (1±rho)/2 across the author-linked contexts and a law-of-one-price book collects rho regardless of truth value. I checked the arithmetic; it is Bayes' rule and cancellation, and it lands. Second, this is the most honestly reported empirical paper in this space I have read in a while: the E-A0 kill criterion fired and they treated it as binding, the E-A3 ablation re-attributes a factor of about 2.4 of the negation reduction to the calibration anchor and says so in the text, and every failed guard (E-A1 task parity, the 7B collapse, lost weights) is on the table, not footnoted away.
What is genuinely new: the bounded-trader-to-relaxation-ladder dictionary giving "degree of understanding" a complexity-theoretic reading with a generalization prediction, the objective-defect result locating the flaw in NLL rather than the architecture, and the three-legged Arbitr scheme. The authors are scrupulous about prior art, exploitability-as-distance, local polytopes, logical induction, and consistency fine-tuning are all cited and disclaimed.
Now the soft spot, and it is load-bearing. Every Inc number and every Arbitr gradient is computed from the first-token answer distribution, and the authors themselves document three silent readout failures plus a numbered-choice format that barely correlated with the other two. The sharper problem is structural: the cross-format book is simultaneously the instrument that caught those failures and the paper's central LOP measurement axis, so the validator and the measured effect are the same tool. During training the model fully controls the quantity being optimized and can satisfy books readout-locally, making first-token answers format-invariant without any underlying credence change. The 0.5-vacuity collapse is the discovered instance of exactly this route; the Brier anchor closes that channel but only on training-time known-truth propositions, while the held-out books behind the headline numbers are unsupervised. The unseen-lexeme transfer (x24 on Phi-3.5) and the cross-family replication are partial evidence against pure readout gaming, but they run inside the same two formats and the same readout, so they do not settle it. A third format, or answer-averaging, or a semantic-clustering readout would.
Also on the table: pre-registration is self-attested dated files, disclosed but with no external deposit, so the dates are unverifiable; no artifacts shipped; and the original Qwen trained-arm adapters are gone, so the direct discrimination test could not run on the main arm. All fixable. The fit-to-metric structure (train on Inc, evaluate on Inc) is real and the E-A3 decomposition mitigates the attribution axis, but the readout-gaming route cuts across both channels regardless.
Verdict: this deserves a serious referee, CONDITIONAL is the right call, not reject and not accept as-is. The theory holds, the claims are scoped, and the honesty means the flaws are findable and addressable. Referees should push for the third-format or independent-readout evaluation, external registration deposit, and artifact release. I would take it to reading group and would cite Thm 7 either way.
The strongest objection
Every Inc number in the paper, from the E-A0 kill verdict to the E-A1b headline reductions, and every Arbitr training gradient, is computed from the first-token answer distribution. The authors document three silent readout failures during the campaign, one of which inverted answers at high coverage, plus a numbered-choice format whose responses only weakly correlated with the others (r about 0.02 to 0.4) ([b233]). After discarding the broken formats, two remain, and the cross-format law-of-one-price book between them is simultaneously (a) the instrument that 'caught' the failures ('all caught by cross-format books rather than the coverage filter', [b233]) and (b) the paper's central LOP axis. The tool that validates the readout is the same tool that measures the effect, which is circular exactly where the claim is strongest. The sharper problem is in training: the trader's profit is computed on the first-token readout, so the model has full control of the quantity being optimized and can satisfy the books readout-locally, by making its first-token answer pattern format-invariant, whether or not its underlying credences changed. The 0.5-vacuity collapse is the discovered instance of this route, and the Brier anchor closes that specific channel, but the anchor only supervises training-time known-truth propositions; the held-out books that yield negation Inc 0.0015 and LOP profit 0.041 are unsupervised and could be satisfied at the readout level. Transfer to 'unseen' patterns runs inside the same two formats and the same readout, so it does not test belief-level coherence either.
The theoretical core (Thm 7, Thm 5, Prop 11) is sound: I re-derived the law-of-one-price book and it collects rho in both worlds, and the convex-relaxation and projection facts are standard and honestly cited. My concern is not with the mathematics but with the empirical link between that mathematics and the measured 'understanding' gains, and there the reader's weakest_assumption is exactly the most load-bearing point. Every quantity that supports a headline claim - the E-A0 kill verdict, the orders-of-magnitude negation reduction, the LOP axis, the transfer results - is computed from the first-token readout, and the paper itself documents that this readout is fragile, with three silent failures and one format that barely correlated with the others. The circularity I emphasize is slightly sharper than the reader's formulation: not just that the measurement presupposes LOP across formats, but that the training objective itself optimizes the same readout it uses to measure success, so the reduced Inc could reflect the model learning to satisfy books at the readout level rather than changing its beliefs. The vacuity collapse to 0.5 is the demonstrated failure mode of exactly this kind, and the anchor closes only the specific 0.5 route, on anchored propositions only. I credit real mitigation: the relation-level selectivity in Table 2 (coherent on capitals, incoherent on mountain heights at 7B) is hard to explain as a uniform readout artifact, the E-A3 training-free baseline behaves sensibly, and the authors disclose every failure. Those mitigate; they do not remove the condition. The concrete test I propose directly closes the gap by measuring the trained model with a readout it never saw during optimization; if the coherence persists, the central empirical claim is genuine. I therefore keep the reader's CONDITIONAL verdict unchanged - the conditions should be read as including independent-readout validation - and I rate my agreement with the reader's weakest_assumption as full.
The referee's report
If the empirical effects survive readout validation, the paper makes a substantial contribution: it turns a philosophical notion of understanding into a graded, trainable, decision-theoretically grounded quantity; Theorem 7 cleanly locates one class of cross-format incoherence in the NLL objective rather than in the architecture; the bounded-trader ladder connects the definition to LP/PSAT complexity; and the E-A3 decomposition plus the scaling-illusion caveat are genuinely informative. The reporting is unusually honest: pre-registration verdicts are given even when they fail, resource gaps are disclosed, and the attribution of the effect to the trader versus the anchor is explicitly decomposed. The central weakness is empirical: every Inc measurement and every Arbitr gradient is computed from first-token readouts, and the paper documents three silent readout failures plus a weakly correlated numbered-choice format. The cross-format books that caught those failures are the same law-of-one-price quantity used as the main measurement axis and as a training target, so the readout is not independently validated.
Pith's reading of the paper (a machine's summary, not the authors')
This paper makes understanding measurable by turning it into a no-arbitrage condition: a model understands a vocabulary to the degree that no computationally bounded trader can lock in guaranteed profit by betting against its stated probabilities on logically related claims, the classical Dutch-book test. Its centerpiece is a theorem that the exact optimum of standard next-token training is itself incoherent: on a corpus written by two internally consistent authors who assert $\varphi$ and $\neg\varphi$ in distinguishable styles, the optimal model prices the claim at $(1+\rho)/2$ in one style and $(1-\rho)/2$ in the other, so a law-of-one-price book collects guaranteed profit $\rho$ regardless of the truth of $\varphi$ and regardless of scale. The paper then introduces Arbitr, a training objective that adds an adversarial trader and a calibration anchor, and reports that it cuts cross-format exploitability by about two orders of magnitude at no measured cost to task accuracy, with transfer to untrained patterns and a second model family. Because full coherence is NP-hard to verify, the paper argues that understanding is inherently graded, and it documents a scaling illusion in which large models register near-zero incoherence while being extremely confident. A sympathetic reading is that if these results hold, cross-phrasing consistency is a trainable property earned by the right objective, not a hidden gift of scale.
What carries the argument. The load-bearing device is the bounded Dutch book: a portfolio supported on at most $k$ propositions instantiating rules of an inference system, whose guaranteed profit is the worst-case minimum over all worlds consistent with the known logical structure, with $T_k$-unexploitability as the definition of degree-$k$ understanding. Two identities carry the argument. Exploitability equals the largest violation of $k$-local coherence constraints, i.e., the $\ell^\infty$ distance to a local Sherali-Adams relaxation of the marginal polytope, which makes bounded trader classes a graded ladder and makes checking full coherence NP-complete. Theorem 7 supplies the closed form: on a two-author corpus, the NLL optimum prices a claim at $(1+\rho)/2$ in the asserting author's style and $(1-\rho)/2$ in the other, so a cross-context book earns exactly $\rho$, the style-author correlation. The training machinery, Arbitr, is a three-legged objective: task loss, an adversarial trader term that penalizes the realized guaranteed profit on batches whose admissible worlds are known by construction, and a Brier calibration anchor that keeps credences informative; Proposition 11, which extends a known projection result, guarantees that coherentizing at any level strictly improves Brier score in every possible world, so the arbitrage term does not trade off against task accuracy at the optimum.
Core claim. On its own terms, the central claim is that the standard next-token objective, not any particular architecture, is what makes language models answer the same question inconsistently across phrasings. Theorem 7 constructs the exact NLL optimum for a corpus in which two internally consistent authors assert $\varphi$ and $\neg\varphi$ with equal weight, with a style correlated with identity; the optimum elicits $q(\varphi;c_1)=(1+\rho)/2$ in one context and $q(\varphi;c_2)=(1-\rho)/2$ in the other, and a portfolio that sells $\varphi$ in the first context and buys it in the second collects guaranteed profit $\rho$ whatever the truth of $\varphi$. Per context the optimum is coherent, so the defect is precisely cross-context: negative log-likelihood never scores one proposition under two elicitation contexts, and the paper's reading is that no NLL optimum can be both a faithful corpus simulator and an elicitation-invariant credence function. The surrounding results set the frame: bounded trader classes match local relaxations of the coherent polytope, so understanding is graded because full coherence is NP-complete to verify, and $\delta$-coherent credences degrade additively along derivation chains, turning overconfidence into a priced arbitrage opportunity. Empirically, deployed models turn out to be exploited exactly on cross-format law-of-one-price books, and training against an adversarial trader with a calibration anchor removes that exploitability by orders of magnitude at zero measured task cost, while the near-zero incoherence seen at 7B parameters often coincides with near-maximal confidence, a confident degeneration the metric cannot by itself distinguish from honest certainty.
The premise it rests on. Every measurement in the paper, including the E-A0 verdict, the training evaluations, and the scaling-illusion reading, rests on credences read from the model's first answer-token, a readout the authors themselves caught failing silently three times with one failure inverting answers at high coverage; if that readout does not track the model's real beliefs, then both the measured incoherence and the training signal could be artifacts of the meter rather than properties of the objective.
If the paper is right (Pith's inferences, not the authors')
Cross-format incoherence is an objective-level fact, not a scale artifact: Theorem 7's arbitrage profit $\rho$ survives at the exact NLL optimum, so only a training term that scores one proposition across contexts can remove it.
Understanding becomes graded and checkable: a model can carry word-level certificates $\mathrm{Inc}_{k,v}$, and training at level $k$ should transfer coherence to unseen instances of patterns at level $k$ while leaving level $k+1$ untouched.
Coherence separates cleanly from truth: a coherent flat-earther is unexploitable by logical books, so coherence is necessary but not sufficient for knowledge, and knowledge enters only through the calibration anchor.
The scaling illusion is a measurement caveat with teeth: at 7B parameters, near-zero measured incoherence and near-maximal confidence arrive together, so exploitability figures must be read alongside an informativeness measure.
My inference: the paper's own documentation of three silent readout failures, one of which inverted answers at high coverage, makes re-eliciting credences by averaging over sampled completions the cheapest decisive check of whether the measured exploitability describes the objective or the meter; the paper itself names this elicitation as its planned foundation.
My inference: the same no-arbitrage definition extends to any declared meaning-preserving transformation, such as translation, modality, or summarization, so law-of-one-price books could certify multilingual or multimodal consistency without retraining the adversarial component.
My inference: Theorem 7's two-author construction is the simplest member of a broader class; adversarially engineering corpora with known author identity would let practitioners audit a deployed model's cross-context arbitrage against a closed-form bound rather than a sampled one.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Does a language model merely predict tokens, or does it understand what it says? We make this question measurable by defining "understanding" through the lens of no-arbitrage. A model understands a vocabulary to a certain degree if a computationally bounded trader cannot extract guaranteed profit by betting against the model's probabilities on logically related claims (a "Dutch book"). We establish three theoretical results: first, because full logical coherence is computationally intractable, understanding is inherently graded, not absolute. Second, we prove that the exact optimum of standard next-token prediction is inherently incoherent across different question formats; the flaw lies in the training objective, not the architecture. Third, we show that uncertainty accumulates predictably along reasoning chains, making unjustified overconfidence an arbitrage opportunity in itself. To address this, we introduce Arbitr, a training framework where an adversarial trader penalizes the model for logical inconsistencies, paired with a calibration anchor to prevent uninformative collapse. Across five pre-registered experiments on Qwen2.5 and Phi-3.5 models, we demonstrate that standard models are highly exploitable across different phrasings. Arbitr reduces this exploitability by orders of magnitude without sacrificing task accuracy, and the effect successfully transfers to unseen logical patterns and new model families. Crucially, we uncover a scaling illusion: at 7B parameters, near-zero measured incoherence often coincides with extreme, unjustified confidence. We conclude that while Arbitr enforces rigorous logical consistency, coherence is a necessary condition for knowledge, but not a sufficient one
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-01
The authors' main claim, as the classifier picked it out 1.00
“We conclude that while Arbitr enforces rigorous logical consistency, coherence is a necessary condition for knowledge, but not a sufficient one”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.90 · human participants 0.07 · controlled comparison 0.87 · randomized assignment 0.04 · states quantitative result 0.58 · compares with prior results 0.90 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Does a language model merely predict tokens, or does it understand what it says? | background 0.84 | 0.04 | 0.02 | unclear 0.97 |
| We make this question measurable by defining "understanding" through the lens of no-arbitrage. | contribution 0.78 | 0.89 | 0.04 | asserted 1.00 |
| A model understands a vocabulary to a certain degree if a computationally bounded trader cannot extract guaranteed profit by betting against the model's probabilities on logically related claims (a "Dutch book"). | contribution 0.73 | 0.32 | 0.06 | asserted 1.00 |
| We establish three theoretical results: first, because full logical coherence is computationally intractable, understanding is inherently graded, not absolute. | result 0.98 | 0.87 | 0.08 | asserted 1.00 |
| Second, we prove that the exact optimum of standard next-token prediction is inherently incoherent across different question formats; the flaw lies in the training objective, not the architecture. | result 0.98 | 0.87 | 0.11 | asserted 1.00 |
| Third, we show that uncertainty accumulates predictably along reasoning chains, making unjustified overconfidence an arbitrage opportunity in itself. | result 0.98 | 0.85 | 0.06 | asserted 1.00 |
| To address this, we introduce Arbitr, a training framework where an adversarial trader penalizes the model for logical inconsistencies, paired with a calibration anchor to prevent uninformative collapse. | method 0.89 | 0.93 | 0.06 | asserted 1.00 |
| Across five pre-registered experiments on Qwen2.5 and Phi-3.5 models, we demonstrate that standard models are highly exploitable across different phrasings. | result 0.99 | 0.95 | 0.06 | asserted 1.00 |
| Arbitr reduces this exploitability by orders of magnitude without sacrificing task accuracy, and the effect successfully transfers to unseen logical patterns and new model families. | result 1.00 | 0.96 | 0.16 | asserted 1.00 |
| Crucially, we uncover a scaling illusion: at 7B parameters, near-zero measured incoherence often coincides with extreme, unjustified confidence. | result 0.96 | 0.94 | 0.08 | asserted 0.97 |
| We conclude that while Arbitr enforces rigorous logical consistency, coherence is a necessary condition for knowledge, but not a sufficient one | interpretation 0.99 | 0.71 | 0.83 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 0.93 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-01: 1 request, 8,288 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
References (13)
- [3] N. Asher, S. Bhar. Strong hallucinations from negation and how to fix them. arXiv:2402.10543, on PithPith's reviewarXiv
- [6] S. K. Manchingal, S. Nikolenko, F. Cuzzolin. Credal large language models for semantic commitment under uncertainty. arXiv:2608.23244, on PithPith's reviewarXiv
- [7] K. Wang et al. CreINNs: credal-set interval neural networks. arXiv:2401.05043, on PithPith's reviewarXiv
- [11] M. A. Mannucci. Logical GANs: adversarial learning through Ehrenfeucht–Fraïssé games. arXiv:2510.22824, on PitharXiv
- [1964] Q. Liu, L. Ye, W. Ma, Y.-C. Chou, A. Yuille. Generative adversarial reasoner: enhancing LLM reasoning with adversarial reinforcement learning. arXiv:2512.16917, on PitharXiv
- [1994] D. Calanzone, S. Teso, A. Vergari. Logically consistent language models via neuro-symbolic integration. arXiv:2409.13724, on PithPith's reviewarXiv
- [1998] S. Garrabrant et al. Logical induction. arXiv:1609.03543, on PithPith's reviewarXiv
- [2000] arXiv:math/0003046. A. Gilio. Probabilistic reasoning under coherence in System P. Ann. Math. AI 34, on PithPith's reviewarXiv
- [2003] I. Pres, L. Ruis, M. Ghebreselassie, B. Z. Li, J. Andreas. Self-CTRL: self-consistency training with reinforcement learning. arXiv:2606.18327, on PithPith's reviewarXiv
- [2022] I. Andrews. Revealed rationality: label-free evaluation and regularization from representation theorems. arXiv:2608.05015, on PithPith's reviewarXiv
- [2023] C. Chen, M. Jörke, A. Goliński, M. Fedzechkina, G. Sapiro, S. Williamson, N. Foti. LLMs are not (consistently) Bayesian: quantifying internal (in)consistencies of LLMs’ probabilistic beliefs. arXiv:2605.06915, on PithPith's reviewarXiv
- [2024] M. Caprio et al. Imprecise Bayesian neural networks. arXiv:2302.09656, on PithPith's reviewarXiv
- [2026] I. Andrews, S. Sarkar. Dutch books for language models. arXiv:2609.02797, on PithPith's reviewarXiv
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment