Pith.

Machine Learning · arXiv:2609.28587 v1 · submitted 23 Sep 2026 · 13 references

NumericJev: Jev-like LLM Numerical Decoding with Multiway Decision Trees

Weiwei Ye, Hangchen Liu, Renhe Jiang

The paper says

“Large language models can interpret natural lan- guage, yet robust decisions remain challenging. Jev-like models expose structured choices, but these interfaces do not directly provide numeri- cal values at a requested precision.”

more of the abstract

We propose NUMERICJEV, a training-free numerical decod- ing algorithm that enables numerical output from any LLM with a Jev-like structured-choice in- terface. Surprisingly, on our arithmetic bench- mark, it outperforms direct selection from a can- didate list containing the correct answer by 2.93 percentage points (Figure 1). Our motivation comes from the observation that numerical range selection is itself a decision problem that Jev- like LLMs can address. NUMERICJEV recur- sively refines a range through a multiway deci- sion tree while retaining the original question in context, without parameter updates or hidden- state access. On a 100-value grid, a ten-way tree requires only two decision rounds. Range- normalized MAE is 1.84% versus 5.18% for di- rect choice. A separate three-date historical- index study yields 4.58% mean relative recall er- ror and 0% readout error when the value is sup- plied. Code is available at https://github. com/Bring-AI/jev-numeric.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 25 Sep 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. A clean, unusually honest methods paper whose formal core is sound, but whose headline outperformance rests on a post-hoc tolerance metric while the primary MAPE endpoint is non-significant and the decisive first-divergence rate is never directly measured. This line is wrong

3 major objections · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 13 references say.

The paper says, at §5.1, Table 1; Abstract (b3)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The abstract and contributions claim a 2.93 percentage point advantage over direct selection, but this number comes from the post-hoc within-5% tolerance metric (§5.3, Figure 1), for which no paired confidence interval is reported. The predeclared primary endpoint, MAPE, shows paired 95% intervals that include zero for every branching factor (b135: [−14.48,7.03], [−16.10,10.31], [−14.05,12.77]).

more of objection 1

The manuscript should either report a paired bootstrap interval for the tolerance difference or temper the abstract claim, since the current phrasing implies a reliable improvement that the primary analysis does not support.

The paper says, at §3.3, Eq. (2); §5

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

Proposition 2 makes the error bound depend on Pr(E_d), the probability that the first incorrect interval selection occurs at depth d. This is the load-bearing reliability quantity, yet the paper never reports estimates of Pr(E_d) or the conditional divergence probabilities, even though Algorithm 1 records the full decision trace.

more of objection 2

Without such estimates, the reader cannot assess whether the method's practical error is dominated by early divergence events (as the bound warns) or whether the model's interval choices are reliable enough. The authors should report Pr(E_1), Pr(E_2), etc., for each branching factor, or explicitly explain why these are unavailable.

This review says the paper's mathematics is wrong. Pith did not check that against the paper's source.

The paper says, at §5.3 and §5.4

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The paper correctly notes that NMAE and MAPE answer different questions, and that near-zero targets dominate MAPE. However, the claimed practical advantage of NumericJev over direct selection is supported only by NMAE (which excludes zero) and by the post-hoc tolerance metric.

more of objection 3

The domain-level results in Table 2 show large reversals (e.g., direct selection has 0.52% MAPE on the ×10⁴ domain vs 30.96% for K=10), indicating that no branching factor is uniformly superior. The central claim of 'improvement over direct choice' therefore appears to be conditional on the loss function and domain. The manuscript should either present a loss-function-agnostic argument for why NMAE is the more relevant metric for the stated goal, or weaken the headline claim.

4 minor objections

The paper says, at Abstract, §5.3

Pith says: minor objection 1 of 4

The abstract's '2.93 percentage points' is not tied to any table or confidence interval; please add a reference to Figure 1 and state that this is a post-hoc tolerance metric, as the paper itself acknowledges in §5.3.

The paper says, at §4, Bootstrap description

Pith says: minor objection 2 of 4

The resampling unit (64 base families) is described only in §4; Figure 1's 'family-bootstrap' is not defined until later. Move the definition earlier for clarity.

show 2 more

The paper says, at Figure 1 caption

Pith says: minor objection 3 of 4

The caption mentions 'hatched bars' but does not explain what the hatching encodes. Please add a legend or explicit note.

The paper says, at §3.4, first sentence

Pith says: minor objection 4 of 4

'Direct selection lists all grid values in one question' – this is clear, but the later claim 'mathematical equivalence does not imply identical model behavior' is important and should be highlighted as a testable hypothesis, not just an assertion.

What would settle it

From the released decision traces for the 256 arithmetic expressions, estimate under the same 64-family bootstrap with seed 20260923 the first-divergence probability Pr(E_1) for K=10 and also K=2,4 in the arithmetic condition, separately by option order and pooled, with 95% percentile-bootstrap intervals. Define a depth-1 error as a selected first-level interval whose half-open description does not contain the verified target. Report these rates together with the W_0 Pr(E_1) contribution to Eq. (2) and compare it with the observed NMAE. If the upper confidence bound for Pr(E_1) is low enough that the depth-1 term is small relative to observed NMAE, the interval-reliability premise is empirically supported; if the rates are high or the interval is wide, the central numerical-readout claim is not established by the current evidence, regardless of aggregate NMAE.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 25 Sep 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 1 min 52 s · 16,335 tokens read, 13,150 written
  3. Editor wrote the verdict above 19 s · 17,105 tokens read, 3,011 written
  4. Referee major revision: 3 major objections, 4 minor 14 s · 16,536 tokens read, 2,189 written

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

You should know two things about this paper. First, it is a genuinely clean methods paper: the training-free multiway interval decoder over a structured-choice interface is well-defined, the depth and error bounds in Propositions 1 and 2 are self-contained and parameter-free, and the prior regression-to-classification reduction is cited honestly (Torgo & Gama 1997; Langford et al. 2006). Second, the headline outperformance claim in the abstract is computed from a post-hoc tolerance metric, while the predeclared primary endpoint — MAPE — shows 95% paired differences against direct choice that include zero for every tree ([−14.48, 7.03] for K=2; b135). That is the main thing to hold in your head before reading further.

The NMAE improvements are real and statistically supported (1.84% vs 5.18%; paired intervals all strictly below zero, b137), so there is a legitimate signal that interval refinement reduces range-normalized error on this benchmark. The paper is unusually honest about scope: it flags the post-hoc metric explicitly in Section 5.3, notes the intervals describe this generator and model version, not a population of tasks (b124), and acknowledges that an early incorrect interval is unrecoverable (b201). The continuous-head baselines are non-comparable — different backbones and prior training tasks — and the authors say so themselves (b205). No attempt to hide the weakness of the primary endpoint.

The stress-test concern holds up. Proposition 2 makes the error depend on first-divergence rates Pr(Ed), and the paper never reports them directly, even though Algorithm 1 records the decision trace and they would be trivially computable from it. The downstream metrics are at best ambiguous evidence for the load-bearing premise that interval choices reliably contain the target. NMAE excludes zero, but NMAE is deliberately insensitive to the near-zero signed-domain errors that dominate MAPE. This is an addressable gap, not a fatal flaw: the method's reliability on this one model is decent in range-normalized terms, just not established by the primary endpoint.

The fix is straightforward and the authors have done most of the hard work already. Report the within-5% tolerance metric as secondary rather than headline, estimate and report Pr(Ed) per depth, and validate interval-selection reliability on additional model versions and task families. None of this requires rethinking the algorithm; the formal core stands.

Who gets value: anyone building numerical readout over hosted choice APIs, plus methodologists who want a good case study in primary-versus-post-hoc endpoint discipline. It deserves a serious referee — the formal parts are sound and the evaluation design is careful enough to engage with properly, even though the empirical support for the headline is thinner than the abstract suggests.

The strongest objection

The central claim — that a Jev-like choice interface can provide numerical readout at user-specified precision — is made true by Proposition 2 only if first-divergence events are rare. Eq. (2) bounds error by ε + Σ W_{d−1} Pr(E_d), and in the 100-cell K=10 design a depth-1 error has W_0 = 100ε, contributing the full interval width. The paper never reports Pr(E_1) or Pr(E_d) directly, even though Algorithm 1 records the decision trace. The downstream metrics are ambiguous as evidence for the needed reliability: the predeclared primary endpoint, MAPE, has 95% paired differences against direct selection that include zero for every tree (§5.1, b135), and the 83.40% within-5% headline is computed on a post-hoc tolerance metric with no reported paired confidence interval. The NMAE differences do exclude zero, but NMAE is range-normalized and deliberately insensitive to the near-zero signed-domain errors that dominate MAPE (§5.4). Thus the load-bearing assumption — that the model's interval choices contain the target often enough for Eq. (2) to be useful — is supported only indirectly. The authors clearly acknowledge early errors can dominate and that the evaluation covers one model and generator (b124, b201, b207), so this is not a scope objection; it is a request to measure the quantity their own bound identifies as decisive.

Proposition 2 is the analytical core of the paper: without estimates of Pr(E_d), the decoder's finite-resolution guarantee is conditional on events that the paper never directly counts. The paper is honest about scope and limitations, and its recorded decision traces, exact-label oracle, and reproducible artifact are strong practices that make the proposed check feasible offline. My concern is not that the method is invalid; it is that the empirical support for the key premise is statistically indirect. The primary MAPE endpoint is non-significant relative to direct selection, while the NMAE and within-5% improvements are either a complementary range-normalized metric or a post-hoc threshold. Measuring Pr(E_1) from logged traces is a decisive and inexpensive check that does not require new model queries. If first-divergence rates are low, the reader's conditional verdict could be upgraded; if not, the abstract's stronger claims should be softened. Because this concern coincides with the reader's weakest_assumption, no verdict change is needed beyond the already-conditional status.

The referee's report

The theoretical contribution is solid: the depth and error bounds are clean, parameter-free, and follow directly from the algorithm's definition. If the method reliably selects the correct interval at shallow depths, it offers a practical way to obtain numerical readouts from choice-only interfaces without training a regression head. However, the empirical evidence for that reliability is thin. The primary endpoint (MAPE) does not distinguish NumericJev from direct selection, the decisive Pr(E_d) quantities are never measured, and the only statistically significant improvement is on NMAE, which is deliberately insensitive to near-zero targets. The headline gain over direct choice rests on a post-hoc tolerance metric with no reported confidence interval. The paper's value therefore depends on whether the missing evidence can be supplied; as it stands, the central claim of outperforming direct selection is not established.

Pith's reading of the paper (a machine's summary, not the authors')

This paper tries to establish that an LLM whose interface only offers categorical choices can still deliver precise numerical values, without any training, parameter updates, or hidden-state access. The idea is to turn numerical readout into a sequence of interval-selection decisions: partition a user-supplied range into $K$ subintervals, let the model pick the interval containing the answer, narrow, and repeat. On 256 arithmetic expressions over a 100-cell grid, ten-way NumericJev puts 83.40% of outputs within 5% relative error and cuts range-normalized mean absolute error from 5.18% (direct selection among 100 listed values) to 1.84%. The authors are explicit that this evidence covers one hosted model version and a generated benchmark, not a population of all numerical tasks.

What carries the argument. The central mechanism is a balanced multiway interval tree (Algorithm 1): at each node the decoder partitions the current index interval $[a,b)$ into $k=\min(K,b-a)$ subintervals via $t_j = a + \lfloor (b-a)j/k \rfloor$, asks the model to choose the interval containing the target, and descends until $b-a=1$. Two propositions carry the argument: Proposition 1 bounds depth by $D=\lceil\log_K N\rceil$ and guarantees exact-oracle readout with error below the resolution $\epsilon$, while Proposition 2 bounds expected error by the width of the parent interval at the first divergence, showing that early mistakes dominate later precision. The external decoder handles all interval arithmetic; the model supplies only contextual judgment about where the answer lies.

Core claim. Numerical readout can be implemented over a structured-choice interface without adding a trained numerical head. NumericJev recursively narrows a supplied range by asking the model to select, at each of at most $\lceil\log_K N\rceil$ steps, which of $K$ subintervals contains the target, then returns the lower endpoint of the final cell. The surprise is that this beats direct selection from a list containing the correct answer by 2.93 percentage points on the within-5%-error tolerance metric, and that its range-normalized-error advantage is statistically robust while the MAPE differences remain within confidence intervals that include zero. The paper stresses that resolution and accuracy are separate properties: many representable values do not imply small error, because an early wrong interval can never be recovered.

The premise it rests on. The whole method rests on the hosted model reliably selecting the interval that actually contains the target, especially at the shallow early decisions, because Proposition 2 shows an early wrong interval dominates the error bound and cannot be repaired by later refinement.

If the paper is right (Pith's inferences, not the authors')

Any Jev-like choice interface can emit arbitrary numerical values at user-specified precision with zero training, making numerical readout a pure decoder problem rather than a modeling problem.

On a 100-cell grid, ten-way refinement produces each output in only two sequential choices, so the method is cheap in rounds regardless of how many distinguishable values the grid holds.

Providing the target value does not make every encoding equivalent: index-digit and bit encodings still fail badly (59.82% and 77.90% MAPE) even when the answer is supplied, so the requested representation matters independently of computation.

Interval trees robustly reduce range-normalized error relative to direct choice, with 95% bootstrap intervals excluding zero, even though their MAPE advantage is not statistically established.

When the value is supplied, ten-way NumericJev reaches 99.80% within-5% accuracy versus 97.27% for direct choice, showing the readout, not the computation, becomes the limiting step.

Because resolution is a caller knob and no model training is involved, the same decoder should port to other Jev-like providers; the natural next test is whether the early-divergence rate, not the fineness of resolution, is what limits accuracy across models.

The historical recall outputs all sit above the reference on three dates, a directional tilt the paper does not explain; probing that bias over many dates could reveal whether it stems from a token prior or a training-data tilt.

The method's dependence on a supplied range and resolution makes it best suited to bounded estimation problems, such as calibration or bounded forecasting, where the containing range is known in advance rather than inferred.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Large language models can interpret natural lan- guage, yet robust decisions remain challenging. Jev-like models expose structured choices, but these interfaces do not directly provide numeri- cal values at a requested precision. We propose NUMERICJEV, a training-free numerical decod- ing algorithm that enables numerical output from any LLM with a Jev-like structured-choice in- terface. Surprisingly, on our arithmetic bench- mark, it outperforms direct selection from a can- didate list containing the correct answer by 2.93 percentage points (Figure 1). Our motivation comes from the observation that numerical range selection is itself a decision problem that Jev- like LLMs can address. NUMERICJEV recur- sively refines a range through a multiway deci- sion tree while retaining the original question in context, without parameter updates or hidden- state access. On a 100-value grid, a ten-way tree requires only two decision rounds. Range- normalized MAE is 1.84% versus 5.18% for di- rect choice. A separate three-date historical- index study yields 4.58% mean relative recall er- ror and 0% readout error when the value is sup- plied. Code is available at https://github. com/Bring-AI/jev-numeric.

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 · 2026-09-25

The authors' main claim, as the classifier picked it out 0.87

“Surprisingly, on our arithmetic bench- mark, it outperforms direct selection from a can- didate list containing the correct answer by 2.93 percentage points (Figure 1).”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.27 · human participants 0.04 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.95 · mentions code or data availability 0.99

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.

SentenceRoleThe authors' own workStates the main conclusionStrength
Large language models can interpret natural lan- guage, yet robust decisions remain challenging. background 0.99 0.05 0.02 asserted 0.93
Jev-like models expose structured choices, but these interfaces do not directly provide numeri- cal values at a requested precision. gap 0.86 0.09 0.03 asserted 1.00
We propose NUMERICJEV, a training-free numerical decod- ing algorithm that enables numerical output from any LLM with a Jev-like structured-choice in- terface. contribution 0.99 0.83 0.32 asserted 1.00
Surprisingly, on our arithmetic bench- mark, it outperforms direct selection from a can- didate list containing the correct answer by 2.93 percentage points (Figure 1). result 1.00 0.95 0.37 asserted 1.00
Our motivation comes from the observation that numerical range selection is itself a decision problem that Jev- like LLMs can address. background 0.48 0.72 0.04 asserted 0.98
NUMERICJEV recur- sively refines a range through a multiway deci- sion tree while retaining the original question in context, without parameter updates or hidden- state access. method 1.00 0.85 0.11 asserted 1.00
On a 100-value grid, a ten-way tree requires only two decision rounds. method 0.98 0.57 0.03 asserted 1.00
Range- normalized MAE is 1.84% versus 5.18% for di- rect choice. result 1.00 0.94 0.15 asserted 1.00
A separate three-date historical- index study yields 4.58% mean relative recall er- ror and 0% readout error when the value is sup- plied. result 1.00 0.89 0.05 asserted 1.00
Code is available at https://github. com/Bring-AI/jev-numeric. contribution 0.73 0.44 0.02 asserted 0.96

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.87

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, 2026-09-25: 1 request, 7,536 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

Figures (4)
Figure 1 of the paper
Figure 1. Numerical output through different interfaces. Percentage of predictions within 5% relative error on all 256 arithmetic expressions; error bars give 95% family-bootstrap intervals. Continuous heads use 4B/9B backbones with LoRA (Hu et al., 2021), trained on structural-causal (SCM) distributions and transferred without arithmetic training; their means are mapped back from [0, 1]. Jev readouts average two option orders. The hatched bars compare direct choice and ten-way NUMERICJEV in the separate supplied-answer condition. This post-hoc threshold metric supplements the primary MAPE…
Figure 2 of the paper
Figure 2. NumericJev as a multiway decision tree. A recorded 10 + 1 query follows four ten-way decisions from [0, 100) to [11, 11.01). Each level partitions only the selected parent; three siblings are displayed and the remaining alternatives are elided. Thick edges show the selected path. The returned lower endpoint is 11.00; the final cell specifies resolution, not confidence. No model parameters are updated. path as a validated distribution over numerical outcomes. 4. Experimental Design Model and access. All choice-interface experi￾ments use the hosted Jev version identified as…
Figure 3 of the paper
Figure 3. Error depends on representation and input condition. Range-normalized absolute error on the fixed affine-grid evaluation. Markers show means and horizontal bars are 95% family-bootstrap intervals. All transforms share 100 cells. NMAE remains defined near zero;
Figure 4 of the paper
Figure 4. A shape repair does not validate uncertainty. Raw and isotonic threshold outputs for the archived expression 7/8 − 1/16, ascending order, first repeat. The reference is the deterministic value 13/16. The displayed repair is computed from the stored probabilities, not from outcome calibration data. H. Reproduction and Artifact Boundaries The manuscript build regenerates tables and vector SVG fig￾ures from the stored metrics. SVG is the source format for every original figure; vector PDF derivatives are included for portable TeX compilation without shell execution. The anonymous manusc

Figures are the authors', reproduced from the paper.

References (13)
  1. [8] URL https://arxiv.org/abs/2405.12856. Torgo, L. and Gama, J. Regression using classification al- gorithms.Intelligent Data Analysis, 1(4):275–292, on PithPith's reviewarXiv
  2. [10] doi: 10.18653/v1/D19-1534. URL https: //aclanthology.org/D19-1534/. Zheng, C., Zhou, H., Meng, F., Zhou, J., and Huang, M. Large language models are not robust multiple choice se- lectors. InInternational Conference on Learning Repre- sentations, on Pithdoi:10.18653/v1/d19-1534
  3. [11] URL https://arxiv.org/abs/ 2309.03882. 8 Numerical Decoding with Multiway Decision Trees A. Proofs and Interpretation of the Error Bound Depth.For an interval with n>K cells, each child has size at most ⌈n/K⌉. For 2≤n≤K , every child is a singleton. Let nd denote the maximum s... on PithPith's reviewarXiv
  4. [100] These quantities mea- sure disagreement, not error against truth; two identical but inaccurate predictions have zero disagreement. E. Exploratory Records and Distinct Protocols The original integer expressions are 1 + 1 , 2 + 3 , 7−4 , 3×4 , 12/3, and 7+8 . Fractional expressi... on Pith
  5. [256] The spline head predicts 49 parameters for a 16-bin monotone rational-quadratic CDF based on the mono- tone spline parameterization of Durkan et al. (2019). The Beta head predicts two shape parameters, constrained above 1.001. All four checkpoints use seed 7, rank-16 LoRA (Hu et al., on Pith
  6. [1997] doi: 10.3233/IDA-1997-1405. Wallace, E., Wang, Y ., Li, S., Singh, S., and Gardner, M. Do NLP models know numbers? probing numer- acy in embeddings. InProceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Con... on Pithdoi:10.3233/ida-1997-1405
  7. [2006] URL https://arxiv.org/abs/1206.6860. Lukasik, M., Narasimhan, H., Menon, A. K., Yu, F., and Kumar, S. Regression-aware inference with LLMs. arXiv:2403.04182, on PithPith's reviewarXiv
  8. [2019] URL https: //arxiv.org/abs/1906.04032. Federal Reserve Bank of St. Louis. S&P 500 daily closing index. FRED series SP500, on PithPith's reviewarXiv
  9. [2021] URL https://arxiv.org/abs/2106.09685. Jev. Choice api documentation. Online documentation, 2026a. Accessed September 23, on PithPith's reviewarXiv
  10. [2022] URLhttps://arxiv.org/abs/2203.02155. Piskorz, J., Kobalczyk, K., and van der Schaar, M. Elicit- ing numerical predictive distributions of LLMs without autoregression. InInternational Conference on Learning Representations, on PithPith's reviewarXiv
  11. [2023] URLhttps://arxiv.org/abs/2310.07820. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. LoRA: Low-rank adaptation of large language models. arXiv:2106.09685, on PithPith's reviewarXiv
  12. [2024] URL https://arxiv. org/abs/2403.04182. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., and... on PithPith's reviewarXiv
  13. [2026] URL https://arxiv.org/ abs/2603.02913. Qwen. Qwen3.5 base model checkpoints. 4B model card; 9B model card, on PitharXiv
Cited by 5 Pith papers

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.

No comments yet.

Sign in to comment