Computation and Language · arXiv:2610.18122 v1 · submitted 13 Oct 2026 · 28 references
Cheap, Consistent, Calibrated, but Biased? Typed Decision Models as Judges
The paper says
“Typed decision models (TDMs) return a probability for each allowed answer instead of generating text, which makes them candidates for cheaply replacing LLM judges that compare two answers. Whether they share the biases of LLM judges is untested.”
more of the abstract
We run the CALM benchmark's bias tests on eight TDMs and seven LLM judges of 2024 and 2026. Every TDM is biased, in three to eight of twelve tests. Emotional tone affects them most, and rewriting the better answer in such a tone lowers the accuracy of a typical TDM from 99% to 42%. TDMs are more biased than 2026 LLM judges. All eight change verdicts when the answers swap places, which the 2026 LLM judges hardly do, and several side with an answer that most people allegedly prefer or that a famous model allegedly wrote. Yet both kinds of judge are tripped up by the same prompts. The TDMs' probabilities show that most tests mainly flip verdicts the TDM was unsure of, and that good calibration does not protect against bias. As TDMs are more consistent and 37 times cheaper than 2026 LLM judges, our results suggest they can be attractive judges if their biases are tested for the task.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 14 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. First systematic bias test of TDMs as judges; rigorous methods, clear results, but the strongest bias claim is partly confounded by GPT-4-Turbo rewrites. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 28 references say.
The paper says, at Abstract and §4.1
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.93):
“Either these items are hard for any model used as a judge, or their rewrites change more than the cue, and only human judgments of the same items could tell the two apart.”
Pith says: Objection 1 of 3 · major
The abstract states 'Emotional tone affects them most' and §4.1 reports that sentiment bias lowers accuracy from 99% to 42%. However, as the authors themselves acknowledge in the Limitations (§5, b180, b223), the sentiment rewrites were generated by GPT-4-Turbo and may alter more than the emotional tone (e.g., clarity and coherence), especially because the fact-related set pairs a clear R1 with a less clear R2.
more of objection 1
The paper states that 'only rewrites that hold clarity fixed could show how much of the sentiment bias is due to tone.' Since sentiment is the largest and most consistent bias, the headline claim requires a qualification in the abstract and conclusion to reflect that part of the observed effect may stem from changes in response quality rather than emotional tone alone.
The paper says, at Abstract and §4.3
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.97):
“Our citations are fake, as in CALM, and some are implausible, so whether punishing them is a bias or sound judgment, and whether the judges also punish real sources, needs a test with real citations.”
Pith says: Objection 2 of 3 · major
Under authority bias, the paper reports that TDMs punish citations (negative shift) and interprets this as bias. The authors note in the Limitations (§5) that the citations are fake and often implausible, and that 'only real citations could show whether punishing them is a bias or sound judgment.' This caveat should be integrated into the main claims about authority bias in the abstract or at least in the results summary, not only in the Limitations section.
The paper says, at §3.3, excess metric
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):
“For an inconsistent judge, E can slightly understate the bias, since a change by chance can undo a change by the cue.”
Pith says: Objection 3 of 3 · major
The excess metric E = CR - RR is defined and justified in Appendix E.1. The derivation assumes that the cue and chance act independently (or that E never overstates bias). The authors acknowledge that E can understate the bias for inconsistent judges.
more of objection 3
For Solar Decide, which has CR = 0.905, the understatement could be nontrivial. The paper uses a smallest effect of interest to mitigate this, but the potential systematic understatement for the least consistent judge should be discussed explicitly when interpreting Solar's borderline results.
6 minor objections
The paper says, at Table 3
Pith says: minor objection 1 of 6
The column 'Layout change (%)' lacks a clear definition in the table caption. It first appears in the main text in §4.4, but readers may not find the description easily. Please add a footnote or clarify in the caption.
The paper says, at §4.2
Pith says: minor objection 2 of 6
The statement 'the 2026 LLM judges differ from those of 2024 mainly in how they treat the presentation of a response' is supported, but the comparison rests on only two 2024 models. A brief caveat would be appropriate.
show 4 more
The paper says, at Appendix D.2
Pith says: minor objection 3 of 6
Figure 3 shows the exact requests for three formats. The long lines are wrapped for display, which is fine, but the figure would benefit from a clearer indication that the CALM format contains the full prompt text as a string, while the native format splits it into fields.
The paper says, at §3.3
Pith says: minor objection 4 of 6
The definition of ECE uses ten equal-width bins of confidence from 0.5 to 1. The paper should note that this binning (as opposed to equal-mass bins) is a standard choice (Guo et al., 2017) but can produce empty bins for some models.
The paper says, at Appendix H.2
Pith says: minor objection 5 of 6
The logistic mixed model in Table 18 reports odds ratios for each test relative to the neutral note. Because the neutral note uses only alignment-set items, the comparisons for sentiment, verbosity, and verbosity (on better) (fact-related set) conflate bias type with dataset. This should be mentioned in the text or table note.
The paper says, at General
Pith says: minor objection 6 of 6
The paper is long but well-structured. The extensive appendices are a strength. A few minor typos: 'shuffling not reported' in Table 4 left column; 'a judge can be biased' is clear but the phrasing could be tightened for consistency.
What would settle it
Re-run the sentiment test using rewrites that explicitly preserve the original response's clarity and coherence (e.g., by instructing GPT-4-Turbo: 'Rewrite the answer to have a {tone} tone without changing the clarity, coherence, or informativeness of the original answer.'). Compare the excess and accuracy drop on the same 150 items. If the excess decreases by more than 5 percentage points (e.g., from median 29% to below 24%), the original estimate is confounded by unintended quality changes. Report the change and a 95% bootstrap interval.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 14 Oct 2026
- Reader holds up
- Skeptic agrees
- Editor wrote the verdict above
- Referee minor revision: 3 major objections, 6 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
This paper does what needed doing: a head-to-head comparison of eight typed decision models (TDMs) and seven LLM judges on the same bias tests from CALM. The answer is clear—every TDM is biased in three to eight of the twelve tests, sentiment and position hit all eight, and TDMs are more biased than 2026 LLM judges but 37× cheaper and far more consistent. The experimental design is a strength: permutation tests with Benjamini–Hochberg correction, bootstrapped intervals, a smallest effect of interest, sensitivity analyses across request formats, and released code and data. The authors are upfront about limitations throughout.
The soft spot is the one the stress-test flags: the headline claim that emotional tone biases all eight TDMs (excess 24–39%) rests on GPT-4-Turbo rewrites that may alter clarity and coherence alongside tone, especially in the fact-related set where the two responses already differ in those dimensions. The paper acknowledges this (b180, b223) and notes that only human judgments controlling for clarity could disentangle the confound. That doesn't break the paper, but it does mean the magnitude of the sentiment effect is less clean than the other biases, which use template-based perturbations (position, bandwagon, compassion fade, etc.) and are unaffected by this confound. The overall conclusion that TDMs are biased is still well-supported; the sentiment result specifically is solid in direction and presence but fuzzy in measured size. A secondary point: the authority bias test uses fake citations, and punishing implausible fake citations may be rational rather than bias—the authors note this too.
Who should read this: anyone using or considering TDMs as cheap judges, especially practitioners who need to know which biases matter most. The guide in Appendix N is genuinely helpful. The paper deserves a thorough peer review; the main revision should be a clearer statement about the sentiment confound's impact on the quantitative claim, possibly with a bound or a sensitivity analysis. I'd bring it to the reading group and would cite it in my own work.
The strongest objection
The paper's central claim that sentiment biases all eight TDMs (Figure 1, excess 24-39%) relies on rewrites generated by GPT-4-Turbo that alter the emotional tone of one response. As the authors note in Limitations (b180, b223), these rewrites may change more than the intended cue: in the fact-related set, the two responses already differ in clarity and coherence by design (Section 3.1). An emotional rewrite that also reduces clarity would make the response legitimately worse, so the observed drop in accuracy (99% to 42%) could partly reflect a rational response to degraded quality rather than a bias toward or against emotional tone. The paper's own statement that 'only rewrites that hold clarity fixed could show how much of the sentiment bias is due to tone' (Limitations) confirms this is an unresolved confound. Since sentiment is the largest and most consistent bias across all TDMs, its interpretation is load-bearing: if the effect is partly an artifact of confounded rewrites, the headline claim that 'emotional tone affects them most' would need qualification. The paper's other bias tests (position, bandwagon, compassion fade, etc.) use template-based perturbations that do not suffer from this confound, so the overall conclusion that TDMs are biased is still well supported. The concern is therefore about the strength and specificity of the sentiment finding, not about the existence of bias per se.
The reader's weakest_assumption correctly identified that GPT-4-Turbo rewrites may alter multiple aspects beyond the intended cue. This is indeed the single most load-bearing concern because sentiment is the strongest and most universal bias. However, the paper acknowledges this limitation explicitly and does not overclaim; its conclusions about other biases (position, bandwagon, etc.) are based on template perturbations that are not confounded. The core finding that every TDM is biased in three to eight tests remains robust even if the sentiment effect is partially explained by quality changes. Therefore, the reader's verdict of ACCEPT should stand unchanged, but the concern warrants noting for future work. The proposed concrete test would resolve whether the sentiment bias is primarily about tone or about confounded quality.
The referee's report
If the results hold, this is the first comprehensive bias evaluation of TDMs as judges, a rapidly growing application. The study is methodologically rigorous (permutation tests with Benjamini–Hochberg correction, bootstrapped intervals, smallest effect of interest, multiple sensitivity analyses). The findings that TDMs share many biases with LLM judges, but are 37× cheaper and more consistent, have clear practical implications for pairwise evaluation pipelines. The paper also provides a practical guide for replacement. The code and data are released, enabling replication and extension.
Pith's reading of the paper (a machine's summary, not the authors')
The paper applies an established set of twelve bias tests—previously used on LLM judges—to eight typed decision models (TDMs), which return probabilities instead of text. Every TDM shows bias in at least three tests, and sentiment (emotional tone) and position (answer order) affect all eight. An emotional rewrite on the correct answer drops a typical TDM's accuracy from 99% to 42%. TDMs are more biased than 2026 LLM judges on position and bandwagon cues, but less biased on verbosity. Despite the biases, TDMs are 37 times cheaper and more consistent, making them attractive judges if their specific biases are tested and mitigated.
What carries the argument. The excess measure $E = CR - RR$, which isolates bias beyond chance by subtracting the consistency rate (verdict changes on repetition) from the robustness rate (verdict changes under perturbation). This separates bias from model inconsistency. For TDMs, the probability shift $\Delta p$ and noise floor $\sigma_0$ further show how cues move the probability without necessarily flipping the verdict.
Core claim. Every TDM is biased in three to eight of twelve tests. Sentiment and position bias all eight models, while five tests bias at most one. The test explains far more of the variance in bias than the model (25% vs. 1.4%). TDMs are more biased than 2026 LLM judges overall: position, bandwagon, and compassion fade bias several TDMs but no 2026 LLM judge. However, both kinds of judge flip on the same prompts (correlation 0.95). The TDMs' probabilities reveal that most cues flip verdicts the TDM was unsure of, but sentiment and authority on the better response flip even confident verdicts. Good calibration does not protect against bias.
The premise it rests on. The rewrites for each bias were generated by GPT-4-Turbo, which may alter multiple aspects of the response beyond the intended cue (e.g., a sentiment rewrite also changes clarity); only human judgments could verify that the rewrites isolate the cue.
If the paper is right (Pith's inferences, not the authors')
TDMs can serve as cheap (37 times cheaper) and consistent judges if their biases are checked for the task.
Position bias must be removed by asking in both orders and averaging the probabilities.
Close calls (probability < 0.7) should be escalated to a second judge, except under sentiment and authority where even confident verdicts flip.
Calibration does not indicate resistance to bias; a well-calibrated TDM can still be strongly biased.
Bias benchmarks should place each cue on both the better and worse response to detect hidden biases that look like robustness.
The strong correlation (0.95) in which prompts flip TDMs and LLM judges suggests that prompt difficulty, not just architecture, drives bias; future work could design bias tests based on prompt features.
The hidden probability shifts (no verdict flip) mean that using TDM probabilities as rewards in RL training may introduce bias even when verdicts appear stable.
The large position bias in TDMs (absent in 2026 LLM judges) may stem from the lack of an explicit explanation step; adding a reasoning head could reduce it.
Because the rewrites were written by another LLM, the measured biases may confound the intended cue with other changes; only human-verified rewrites can isolate the true cue.
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
Typed decision models (TDMs) return a probability for each allowed answer instead of generating text, which makes them candidates for cheaply replacing LLM judges that compare two answers. Whether they share the biases of LLM judges is untested. We run the CALM benchmark's bias tests on eight TDMs and seven LLM judges of 2024 and 2026. Every TDM is biased, in three to eight of twelve tests. Emotional tone affects them most, and rewriting the better answer in such a tone lowers the accuracy of a typical TDM from 99% to 42%. TDMs are more biased than 2026 LLM judges. All eight change verdicts when the answers swap places, which the 2026 LLM judges hardly do, and several side with an answer that most people allegedly prefer or that a famous model allegedly wrote. Yet both kinds of judge are tripped up by the same prompts. The TDMs' probabilities show that most tests mainly flip verdicts the TDM was unsure of, and that good calibration does not protect against bias. As TDMs are more consistent and 37 times cheaper than 2026 LLM judges, our results suggest they can be attractive judges if their biases are tested for the task.
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 · 2026-10-14
The authors' main claim, as the classifier picked it out 1.00
“As TDMs are more consistent and 37 times cheaper than 2026 LLM judges, our results suggest they can be attractive judges if their biases are tested for the task.”
How the main claim is stated: hedged possibility 1.00
What the abstract itself states, as probabilities: claims first 0.13 · states limitation 0.86 · human participants 0.13 · controlled comparison 0.81 · randomized assignment 0.03 · states quantitative result 0.99 · compares with prior results 0.94 · mentions code or data availability 0.03
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 |
|---|---|---|---|---|
| Typed decision models (TDMs) return a probability for each allowed answer instead of generating text, which makes them candidates for cheaply replacing LLM judges that compare two answers. | background 1.00 | 0.07 | 0.03 | hedged possibility 0.84 |
| Whether they share the biases of LLM judges is untested. | gap 1.00 | 0.05 | 0.02 | asserted 0.67 |
| We run the CALM benchmark's bias tests on eight TDMs and seven LLM judges of 2024 and 2026. | method 0.74 | 0.95 | 0.03 | asserted 0.98 |
| Every TDM is biased, in three to eight of twelve tests. | result 1.00 | 0.93 | 0.21 | asserted 1.00 |
| Emotional tone affects them most, and rewriting the better answer in such a tone lowers the accuracy of a typical TDM from 99% to 42%. | result 1.00 | 0.92 | 0.11 | asserted 1.00 |
| TDMs are more biased than 2026 LLM judges. | result 1.00 | 0.92 | 0.15 | asserted 1.00 |
| All eight change verdicts when the answers swap places, which the 2026 LLM judges hardly do, and several side with an answer that most people allegedly prefer or that a famous model allegedly wrote. | result 1.00 | 0.90 | 0.08 | asserted 0.99 |
| Yet both kinds of judge are tripped up by the same prompts. | result 0.98 | 0.83 | 0.05 | asserted 1.00 |
| The TDMs' probabilities show that most tests mainly flip verdicts the TDM was unsure of, and that good calibration does not protect against bias. | result 0.75 | 0.94 | 0.16 | asserted 0.99 |
| As TDMs are more consistent and 37 times cheaper than 2026 LLM judges, our results suggest they can be attractive judges if their biases are tested for the task. | interpretation 0.97 | 0.64 | 0.80 | hedged possibility 1.00 |
What kind of paper the classifier reports this is: empirical study 1.00 · 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, 2026-10-14: 1 request, 7,497 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 (7)
Figures are the authors', reproduced from the paper.
References (28)
- [1] Identifying the subject and purpose: The sentence begins by discussing the author's intention ("Buradaki amacım"), indicating what they set out to do. This is translated to reflect the aim of writing about a specific topic, in this case, the Serbian national team on Pith
- [2] Expressing restriction: The phrase "bu yüzden de" implies a restriction or a consequence of the previous statement, which is translated as "therefore" to indicate the cause-and-effect relationship between the purpose mentioned and the subsequent action taken on Pith
- [3] Describing the action taken: "konseri ... izlemekten ba¸ ska seçene˘gim yoktu" expresses a lack of alternatives to a specific action. The phrase is translated as "I had no choice but to watch the concert" to reflect this lack of alternatives on Pith
- [4] Indicating location: The phrase "otelimizin barından" specifies the location from which the action took place. It is important to translate this to show where the author ended up watching the concert ("from the bar of our hotel"), as it adds context to the action and the limit... on Pith
- [5] LLM evaluators recognize and favor their own generations. InAdvances in Neural Information Processing Systems 37, pages 68772–68802. Perplexity. 2026. Perplexity decider v1.1 27b. https://huggingface.co/perplexity-ai/ pplx-decider-v1.1-27b . Model card on Hugging Face, accesse... on Pith
- [6] Choose the model and the request format for the task.No TDM is the least biased in every test, so pick the one whose weak spots matter least for the inputs (Appendix L). The request format changes the bias of most models little, but moving the same text into another format cha... on Pith
- [7] JudgeLM: Fine-tuned large language models are scalable judges. InThe Thirteenth International Conference on Learning Representations. 12 A Differences from CALM CALM This paper Judge and request Judges six LLMs at temperature 0.7 eight TDMs, and seven LLM judges on the same pr... on Pith
- [8] in the CALM format for the margins δ= 0.02 , 0.05 and 0.10, the TDMs in the native (nat.) and decomposed (dec.) formats at δ= 0.05 , and the LLM judges of 2026 (of 5) and 2024 (of 2) atδ= 0.05. Averaged over the twelve tests, the median TDM changes 11.5 of 100 verdicts beyond ... on Pith
- [12] Using past tense: The original sentence is in the past tense, which needs to be retained in the translation to reflect that the events have already taken place. R2 (rejected) First, I'll read and understand the text given by the user. Then, using my natural language processing... on Pith
- [13] Calculate the interquartile range (IQR) of the dataset. The IQR is the difference between the 75th percentile (Q3) and the 25th percentile (Q1) of the data on Pith
- [14] Define a threshold as a multiple of the IQR. Any value outside of this threshold is considered an outlier on Pith
- [15] Remove the outliers from the dataset. Here's an example Python code to achieve this: ```python import numpy as np def remove_outliers(data, threshold=1.5): q1 = np.percentile(data, 25) q3 = np.percentile(data, 75) iqr = q3 - q1 lower_bound = q1 - threshold * iqr upper_bound = ... on Pith
- [16] Keep the factual content and expression style consistent with the original answer on Pith
- [17] Please modify the reasoning in the answer above to be completely chaotic and incorrect on Pith
- [18] Ensure that the conclusion of the answer remains unchanged on Pith
- [19] In your response, do not explain why the reasoning is incorrect or identify the errors. Example: GSM8K, broken reasoning Question A magazine costs half as much as a book. The book costs $4. A pen costs $1 less than a magazine. How much is the pen? R1 with broken reasoning In t... on Pith
- [21] Ask in both orders.Every TDM we tested has a position bias, which no 2026 LLM judge has. Judging each pair in both orders and averaging the two probabilities removes it. This doubles the cost and changed the accuracy of our TDMs by−0.2to+3.8points (Appendix M) on Pith
- [22] Remove cues the task does not need.Strip claimed majorities, popularity counts and the names of models from the input before judging on Pith
- [23] Do not expect a TDM to handle tone and citations better.Both kinds of judge punish an emotional tone and a fake citation even on the better response. Under authority (on better), the accuracy of the median TDM falls from 0.66 to 0.41, and under a tone on the better response fr... on Pith
- [24] Treat close calls as uncertain.Outside sen- timent and authority (on better), three in five flipped verdicts come from verdicts given with a probability below 0.7, which are about a quarter of all verdicts (Figure 2c, Appendix H). Such verdicts can be passed on to a second jud... on Pith
show 8 more
- [25] Use the probabilities with care.Calibration differs widely between the TDMs and does not go with less bias (Appendix I). A cue can also move the probability by more than 0.10 without flipping the verdict, for the median TDM on 19 to 24% of the prompts under authority, position... on Pith
- [27] Read the interface.Check what the model returns on a 50/50 answer and how it defines its confidence. Three of our models always return the first option on a 50/50 answer (Appendix L) on Pith
- [28] Test again after an update.TDMs are new and change quickly. With our released code, a full bias test of one TDM in the CALM or native format costs US$0.26 to 3.11 (Appendix J). Whoever tests a judge, TDM or LLM, should look beyond the robustness rate. It mixes bias with chance... on Pith
- [2021] Training verifiers to solve math word prob- lems.arXiv preprint arXiv:2110.14168. Steven Denney and Matthew DiGiuseppe. 2026. JEV versus LLMs: Accuracy, cost and calibration on seven political science replications.arXiv preprint arXiv:2610.06625. Tobias Deußer, Lorenz Sparrenb... on PithPith's reviewarXiv
- [2023] Direct preference optimization: Your language model is secretly a reward model. InAdvances in Neural Information Processing Systems, volume 36. Amir Rafe and Subasish Das. 2026. Benchmarking Sys- tem One decision models against trained classifiers and language models for autom... on PithPith's reviewarXiv
- [2024] LLMs-as-judges: A comprehensive survey on LLM-based evaluation methods.arXiv preprint arXiv:2412.05579. Jiawei Li. 2026. Fast models, slow evidence: A paired and self-audited evaluation of System-1 decision models for LLM agent harnesses.arXiv preprint arXiv:2610.02267. Yubo L... on PithPith's reviewarXiv
- [2025] An empirical study of LLM-as-a-judge for LLM evaluation: Fine-tuned judge model is not a general substitute for GPT-4. InFindings of the As- sociation for Computational Linguistics: ACL 2025, pages 5880–5895, Vienna, Austria. Association for Computational Linguistics. Jiaju Hu... on PithPith's reviewarXiv
- [2026] Evaluating and benchmarking the System One model Jev.arXiv preprint arXiv:2609.37647. Yann Dubois, Balázs Galambosi, Percy Liang, and Tat- sunori B. Hashimoto. 2024. Length-controlled Al- pacaEval: A simple way to debias automatic evalua- tors.arXiv preprint arXiv:2404.04475. ... on PithPith's reviewarXiv
Cited by 4 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow cs.AI · Nov 2026
- Understanding Jev-Style System One Models: Does Structure Reduce Selection Bias? cs.CL · Oct 2026
- INVAR-Agri: Detecting Discrepancies in Multilingual Operator Manuals of Agricultural Implements with a Fine-Tuned LLM cs.CL · Oct 2026
- VisualJevBench: A Benchmark and Black-Box Attacks for Visual Typed Decision Models cs.CR · Nov 2026
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