Pith. sign in

REVIEW 5 major objections 4 minor 24 references

The paper claims that symbolic input features—especially modifiers and named entities—reliably trigger hallucination in Gemma models at every scale, with rates still near 64–79% even at 27B parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 22:13 UTC pith:RUDXGU5W

load-bearing objection A promising question undermined by a missing control, an undescribed annotation protocol, and an internal contradiction; the per-property rates are co-occurrence, not causation. the 5 major comments →

arxiv 2509.09715 v1 pith:RUDXGU5W submitted 2025-09-09 cs.CL cs.AI

Investigating Symbolic Triggers of Hallucination in Gemma Models Across HaluEval and TruthfulQA

Naveen Lamba , Sanju Tiwari , Manas Gaur This is my paper
classification cs.CL cs.AI
keywords hallucinationsymbolic triggersGemma-2model scalingprompt formatsattention analysisHaluEvalTruthfulQA
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that a large share of LLM hallucination traces to specific symbolic prompt features, not just model size or task difficulty. By converting QA pairs from HaluEval and TruthfulQA into QA, multiple-choice, and odd-one-out formats, the authors isolate five symbolic properties—modifiers, named entities, numbers, negation, and exceptions. They report hallucination rates of 79.0% at 2B, 73.6% at 9B, and 63.9% at 27B, with modifiers and named entities reaching 84–95% and 84–94% respectively. Because these rates drop only modestly with scale and persist across task formats, the authors conclude that current architectures have a structural weakness in encoding and reasoning over symbolic structure. If correct, this redirects hallucination mitigation from scaling and prompting toward targeted symbolic-reasoning mechanisms.

Core claim

The paper's central claim is that hallucination in the Gemma-2 family is consistently provoked by the presence of symbolic linguistic properties in the input. Using 600 transformed prompt instances (100 per dataset per format), the authors find that modifiers (adjectives, adverbs, verbs) and named entities are the strongest triggers, with per-property hallucination rates exceeding 84% and 83% respectively across all model sizes. Scaling from 2B to 27B reduces the overall hallucination rate by only about 15 percentage points, and the same properties remain the most problematic. Attention analysis at mid-to-deep layers shows lower attention allocated to symbolic tokens in MCQ formats, which co

What carries the argument

The central object is the symbolic property tag set: modifiers, named entities, numbers, negation, and exceptions, identified in each prompt and then used to compute per-property hallucination rates. The method converts each QA pair into three task formats (QA, MCQ, Odd-One-Out) to test whether triggers generalize; attention scores to symbolic tokens at selected layers (e.g., Layers 10/20 for 2B) and token-length binning are then used to connect surface hallucinations to internal processing. Together these allow the authors to attribute hallucination rates to input symbolic structure rather than to dataset or format artifact.

Load-bearing premise

The paper assumes that tagging a prompt for a symbolic property and then computing the hallucination rate among tagged prompts isolates that property's triggering effect; because the tags overlap and cover most prompts, the per-property rates may just mirror the baseline hallucination rate of the sampled subset.

What would settle it

A matched ablation would settle it: take the 100 HaluEval and 100 TruthfulQA prompts, strip or neutralize one symbolic property at a time (e.g., paraphrase away modifiers), and re-run the same Gemma models; if hallucination rates do not drop when the property is removed, the trigger claim collapses. A simpler check is to compute the hallucination rate on prompts that contain no tagged property; if it is roughly equal to the per-property rates, the properties are not doing causal work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Larger model scale alone will not eliminate hallucination: the overall rate only drops 15 percentage points from 2B to 27B, and the same symbolic properties dominate at every size.
  • Open-ended QA is the most vulnerable format; giving the model constrained choices (MCQ) or asking it to pick an odd-one-out reduces, but does not remove, the trigger effect.
  • Prompts rich in modifiers and named entities are the highest-risk inputs, so systems that must be reliable should verify or ground those tokens before trusting the answer.
  • Attention to symbolic tokens at mid-to-deep layers correlates with hallucination: lower attention to these tokens goes with more frequent hallucination, making attention a possible early warning signal.
  • The trigger effect is non-monotonic with length: the 10–30 token window is the most dangerous for modifier- and entity-driven hallucination, while longer contexts are somewhat safer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the property definitions are broad (modifiers include most adjectives, adverbs, and verbs; named entities appear in nearly every prompt), the paper's per-property rates may partly reflect the overall baseline hallucination rate of the 100-sample subset rather than a unique effect of that property. A cleaner design would match prompts on all other properties and test each one's marginal con
  • If the causal story holds, then interventions that boost attention to symbolic tokens at the identified layers—such as contrastive decoding or targeted prompting—should cut hallucination more than scaling would; this is a testable prediction the paper does not make.
  • The same transformation pipeline could be applied to other model families and other hallucination benchmarks; the paper lists LLaMA, Mistral, and GPT as future work, but a re-analysis of HaluEval and TruthfulQA with balanced property subsets would be the quickest direct test.
  • The paper's 100-sample subsets and manual annotation mean the headline percentages have sizable margins of error; replicating on the full datasets or with automated NER and modifier tagging would show whether the rankings (modifiers > named entities > numbers > negation/exceptions) survive.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper studies whether five 'symbolic' linguistic properties (modifiers, named entities, numbers, negation, exceptions) act as triggers of hallucination in Gemma-2 2B/9B/27B. It converts 100 samples each from HaluEval and TruthfulQA into QA, MCQ, and Odd-One-Out formats (600 instances), computes per-property hallucination rates, and reports attention scores to symbolic tokens at selected layers and token-length-binned hallucination rates. The authors report overall hallucination rates of 79.0%, 73.6%, and 63.9% for the three scales and conclude that symbolic properties, especially modifiers and named entities, reveal a scale-independent fundamental weakness in LLM processing.

Significance. The question is important: if some input properties reliably trigger hallucination across scales and formats, that would be a valuable diagnostic and interpretability target. The paper has sensible raw materials: two established benchmarks, three members of one model family, a format-conversion design, and attention analyses at model-dependent layers. However, the current evidence does not support the causal claim. There is no property-free baseline, the hallucination-annotation procedure is unspecified, the per-property rates are conditional on property presence rather than causal contrasts, and the attention/length analyses lack defined aggregation and inferential statistics. The paper therefore reads as a correlational report whose central, stronger conclusions are not established. The strengths are real but do not compensate for the missing control.

major comments (5)
  1. [§3.1, §3.3, Table 1] The central causal claim is not testable from the reported design. All 600 transformed instances are 'verified to contain one or more target symbolic properties' (§3.1), and the per-property hallucination rate is the proportion of hallucinated outputs among instances containing that property (§3.3). There is no matched property-free control, and no random baseline from the original datasets is reported. Because modifiers are defined to include adjectives, adverbs, and verbs and named entities are common, the high rates in Table 1 (e.g., modifiers 84.76% on HaluEval/2B) are exactly what would be expected if the datasets simply have a high base hallucination rate. The abstract's language ('reason for the hallucinations', 'continue to confuse the models') is therefore unsupported.
  2. [§3.3, Tables 1 and 3] The hallucination label is not operationalized. 'A prediction was marked as a hallucination if it was factually incorrect' appears without specifying whether labels came from human annotators, automatic metrics, or the dataset keys; no agreement or quality control is reported. No sample sizes per property or per token-length bin, confidence intervals, or significance tests are given. This matters for specific cells such as HaluEval Exceptions at 100.0% (Table 1, QA 2B) and the 0%/100% entries in Table 3, which may be based on a handful of items and make the cross-property rankings unreliable.
  3. [§4.2, Table 2] The paper claims that 'MCQ prompts result in consistently higher hallucination frequency than QA' and that 'QA format produces the highest hallucination rates', but no hallucination-rate table for MCQ/OOO is presented; Table 2 contains only attention values. In addition, the attention aggregation is undefined: it is not stated how symbolic tokens were selected, how attention over heads/layers/tokens was averaged, or how the reported numbers are normalized across prompts of different lengths. Without these definitions and a correlation between attention and hallucination, the claim that lower symbolic attention 'correlates with' higher hallucination is unsupported.
  4. [§4.3, Table 3] The 'activation-level traces' section reports no activation traces; it reports token-length-binned hallucination percentages. The non-monotonic-length conclusion is not supported by the table: several bins have 0% by construction ('absence of the corresponding symbolic property') and some have 100% (e.g., TruthfulQA 50+ for 9B), with no counts. The statement that rates peak at 10–30 tokens and decline for 40+ is contradicted by entries such as Modifiers 66.67% (40–49) and 100% (50+ TruthfulQA 9B). A proper analysis would report counts, confidence intervals, and a regression or at least trend test.
  5. [§3.2, §3.1] The property taxonomy is too broad to isolate individual triggers. 'Modifiers' are defined as adjectives, adverbs, and verbs—effectively most content words—so nearly every prompt contains them; 'named entities' are highly frequent in both benchmarks; and exceptions are defined functionally rather than syntactically. The paper does not report inter-annotator agreement or the distribution of properties across the 600 prompts. Since instances can contain multiple properties, the per-property rates are not independent and cannot be interpreted as isolated symbolic effects. The conversion step also assumes that QA-to-MCQ/OOO reformatting preserves the target property, but this is not validated.
minor comments (4)
  1. [Abstract, §1, Figure 3] Minor formatting and notation issues: 'LLMs(LLMs)' missing space, 'Gemma-2-2b' vs 'Gemma-2-2B' inconsistency, and the figure caption uses a different casing from the rest of the paper.
  2. [Appendix A] The 'GitHub' entry in Online Resources is a bare string, not a URL; code and data are not accessible, which prevents replication.
  3. [§4.2] The phrase 'as seen in prior hallucination rate and effect metrics' is a dangling pointer; no table or section containing MCQ/OOO hallucination rates exists in the manuscript.
  4. [References] Several references are incomplete or malformed, e.g., [4] has 'Conference acronym XX' and a placeholder DOI, and [23] is extremely long with inconsistent formatting. A cleaned reference list is needed.

Circularity Check

0 steps flagged

No circular derivation: property tags and hallucination labels are assigned independently; the reported percentages are measurements, not fitted predictions.

full rationale

The paper does not derive its headline rates from a fitted parameter or from a definition that presupposes the conclusion. Section 3.1 converts samples from HaluEval and TruthfulQA, Section 3.2 defines the five symbolic property categories, and Section 3.3 computes each property-specific rate as (# hallucinated instances containing the property)/(# instances containing the property). Hallucination is annotated independently as 'factually incorrect' output, not as 'contains a symbolic property,' so the numerator and denominator are not the same quantity. No load-bearing self-citation occurs: the citations to prior work with overlapping authorship (e.g., [6], [21]) are contextual and do not supply the empirical claim; the layer-selection choice is attributed to an external source [24]. The absence of a property-free control set is a real validity limitation for the causal claim that these properties 'trigger' hallucination, because every transformed instance was deliberately selected to contain at least one target property (Section 3.1), so conditional rates can track the dataset-wide hallucination base rate. That is a confound, not a circular reduction: the paper's own equations do not make the reported rates equal to an input by construction. Attention-score interpretation is post-hoc and correlational, but again not a self-referential derivation. Under the hard rule that circularity requires an exhibited reduction, no significant circularity is found.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The paper's central claim rests on domain assumptions about how hallucination is judged, how prompt transformations preserve symbolic content, and how attention layers were selected. No formal derivation or fitted parameters are involved.

axioms (4)
  • domain assumption Hallucination is operationalized as factually incorrect responses; the paper does not specify the evaluation procedure (human annotation, model judge, or metric).
    Section 3.3 states 'A prediction was marked as a hallucination if it was factually incorrect' with no method described.
  • ad hoc to paper Converting QA pairs to MCQ and OOO formats preserves symbolic properties and isolates their effect; hallucinations can be attributed to symbolic processing rather than prompt ambiguity.
    Section 3.1: 'Prompt design was carefully managed so that hallucinations, when they occur, can be attributed to model reasoning and symbolic processing rather than prompt ambiguity.'
  • domain assumption Attention at mid/deeper layers selected per Wu et al. captures symbolic processing in Gemma models.
    Section 3.3: 'Following Wu et al.'s approach, which emphasizes mid-to-deeper layers where semantic integration peaks.'
  • domain assumption The 100-sample subsets are representative and each symbolic property is adequately represented.
    Section 3.1 says the samples were 'verified to contain one or more target symbolic properties' but provides no balance or prevalence statistics.

pith-pipeline@v1.3.0-alltime-deepseek · 12613 in / 13845 out tokens · 135795 ms · 2026-08-04T22:13:03.121349+00:00 · methodology

0 comments
read the original abstract

Hallucination in Large Language Models (LLMs) is a well studied problem. However, the properties that make LLM intrinsically vulnerable to hallucinations have not been identified and studied. This research identifies and characterizes the key properties, allowing us to pinpoint vulnerabilities within the model's internal mechanisms. To solidify on these properties, we utilized two established datasets, HaluEval and TruthfulQA and convert their existing format of question answering into various other formats to narrow down these properties as the reason for the hallucinations. Our findings reveal that hallucination percentages across symbolic properties are notably high for Gemma-2-2B, averaging 79.0% across tasks and datasets. With increased model scale, hallucination drops to 73.6% for Gemma-2-9B and 63.9% for Gemma-2-27B, reflecting a 15 percentage point reduction overall. Although the hallucination rate decreases as the model size increases, a substantial amount of hallucination caused by symbolic properties still persists. This is especially evident for modifiers (ranging from 84.76% to 94.98%) and named entities (ranging from 83.87% to 93.96%) across all Gemma models and both datasets. These findings indicate that symbolic elements continue to confuse the models, pointing to a fundamental weakness in how these LLMs process such inputs--regardless of their scale.

Figures

Figures reproduced from arXiv: 2509.09715 by Manas Gaur, Naveen Lamba, Sanju Tiwari.

Figure 1
Figure 1. Figure 1: Examples of symbolic triggers causing hallucination across all Gemma model sizes (2B, 9B, 27B). Color coding: blue = modifier, red = negation, purple = exception, green = named entity, yellow = number. This paper addresses this gap by identifying and describing symbolic and interpretable knowledge properties that reliably trigger hallucination across natural language understanding task types and model scal… view at source ↗
Figure 2
Figure 2. Figure 2: Methodology followed for the hallucination analysis in this research structurally complex or logically nuanced inputs. 3. Methodology Our approach involved taking existing datasets, converting them into different question formats, and then testing how three different sizes of Gemma models (2B, 9B, and 27B) responded to questions containing specific symbolic triggers like modifiers, numbers, and named entit… view at source ↗
Figure 3
Figure 3. Figure 3: Examples of questions and answers in different formats(QA, MCQ and Odd-One-Out), generated from Gemma-2-2b. your reasoning." Odd One Out Prompt: "Identify the item that does not belong in the list. Explain your reasoning." For example: • QA Prompt: "Answer the following question in one short, factual sentence." • MCQ Prompt: "Consider the following multiple-choice question. Pick the correct answer and expl… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 7 canonical work pages · 5 internal anchors

  1. [1]

    Kamalloo, N

    E. Kamalloo, N. Dziri, C. L. Clarke, D. Rafiei, Evaluating open-domain question answering in the era of large language models, Proceedings of the Annual Meeting of the Association for Computational Linguistics 1 (2023) 5591–5606. URL: https://arxiv.org/pdf/2305.06984. doi:10.18653/v1/2023. acl-long.307

    Pith/arXiv arXiv 2023

  2. [2]

    D. V. Veen, C. V. Uden, L. Blankemeier, J.-B. Delbrouck, A. Aali, C. Bluethgen, A. Pareek, M. Polacin, E. P. Reis, A. Seehofnerová, N. Rohatgi, P. Hosamani, W. Collins, N. Ahuja, C. P. Langlotz, J. Hom, S. Gatidis, J. Pauly, A. S. Chaudhari, Clinical text summarization: Adapting large language models can outperform human experts, Research Square (2023) rs...

    doi:10.21203/rs.3.rs-3483777/v1 2023

  3. [3]

    Yugeswardeenoo, K

    D. Yugeswardeenoo, K. Zhu, S. O’Brien, Question-analysis prompting improves llm performance in reasoning tasks (2024). URL: https://arxiv.org/pdf/2407.03624

    Pith/arXiv arXiv 2024

  4. [4]

    S. Guan, H. Xiong, J. Wang, J. Bian, B. Zhu, J. guang Lou, Evaluating llm-based agents for multi-turn conversations: A survey, Proceedings of Make sure to enter the correct conference title from your rights confirmation emai (Conference acronym ’XX) 1 (2025). URL: https://arxiv.org/pdf/2503.22458. doi:XXXXXXX.XXXXXXX

    arXiv 2025

  5. [5]

    Maynez, S

    J. Maynez, S. Narayan, B. Bohnet, R. McDonald, On faithfulness and factuality in abstractive summa- rization, Proceedings of the Annual Meeting of the Association for Computational Linguistics (2020) 1906–1919. URL: https://arxiv.org/pdf/2005.00661. doi:10.18653/v1/2020.acl-main.173

    Pith/arXiv arXiv 2020

  6. [6]

    Govil, H

    P. Govil, H. Jain, V. Bonagiri, A. Chadha, P. Kumaraguru, M. Gaur, S. Dey, Cobias: Assessing the contextual reliability of bias benchmarks for language models, in: Proceedings of the 17th ACM Web Science Conference 2025, 2025, pp. 460–471

    work page 2025

  7. [7]

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. Bang, D. Chen, W. Dai, H. S. Chan, A. Madotto, P. Fung, Survey of hallucination in natural language generation, ACM Computing Surveys 55 (2023) 1–38. URL: http://arxiv.org/abs/2202.03629. doi:10.1145/3571730, arXiv:2202.03629 [cs]

    Pith/arXiv arXiv 2023

  8. [8]

    Huang, X

    L. Huang, X. Feng, B. Qin, T. Liu, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, ACM Transactions on Information Systems 1 (2024). doi:10.1145/3703155

    doi:10.1145/3703155 2024

  9. [9]

    Y. Sun, Z. Yin, Q. Guo, J. Wu, X. Qiu, H. Zhao, Benchmarking hallucination in large language models based on unanswerable math word problem (2024). URL: https://arxiv.org/pdf/2403.03558

    Pith/arXiv arXiv 2024

  10. [10]

    A Survey on Symbolic Knowledge Distillation of Large Language Models

    K. Acharya, A. Velasquez, H. H. Song, A survey on symbolic knowledge distillation of large language models, IEEE Transactions on Artificial Intelligence 5 (2024) 5928–5948. URL: http: //arxiv.org/abs/2408.10210http://dx.doi.org/10.1109/TAI.2024.3428519. doi:10.1109/TAI.2024. 3428519

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Weston, A

    J. Weston, A. Bordes, S. Chopra, A. M. Rush, B. V. Merriënboer, A. Joulin, T. Mikolov, Towards ai-complete question answering: A set of prerequisite toy tasks, 4th International Conference on Learning Representations, ICLR 2016 - Conference Track Proceedings (2015). URL: https: //arxiv.org/pdf/1502.05698

    Pith/arXiv arXiv 2016

  12. [12]

    Zhang, Y

    Y. Zhang, Y. Li, L. Cui, D. Cai, L. Liu, T. Fu, X. Huang, E. Zhao, Y. Zhang, Y. Chen, L. Wang, A. T. Luu, W. Bi, F. Shi, S. Shi, Siren’s song in the ai ocean: A survey on hallucination in large lan- guage models, 2023. URL: http://arxiv.org/abs/2309.01219. doi:10.48550/arXiv.2309.01219, arXiv:2309.01219 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.01219 2023

  13. [13]

    Z. Zhao, S. B. Cohen, B. Webber, Reducing quantity hallucinations in abstractive summarization, Findings of the Association for Computational Linguistics Findings of ACL: EMNLP 2020 (2020) 2237–2249. URL: https://arxiv.org/pdf/2009.13312. doi:10.18653/v1/2020.findings-emnlp. 203

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  14. [14]

    Durmus, H

    E. Durmus, H. He, M. Diab, Feqa: A question answering evaluation framework for faithfulness assessment in abstractive summarization, Proceedings of the Annual Meeting of the Association for Computational Linguistics (2020) 5055–5070. URL: http://arxiv.org/abs/2005.03754http://dx.doi. org/10.18653/v1/2020.acl-main.454. doi:10.18653/v1/2020.acl-main.454

    Pith/arXiv arXiv 2020

  15. [15]

    Yao, K.-P

    J.-Y. Yao, K.-P. Ning, Z.-H. Liu, M.-N. Ning, Y.-Y. Liu, L. Yuan, Llm lies: Hallucinations are not bugs, but features as adversarial examples (2023). URL: https://arxiv.org/pdf/2310.01469

    Pith/arXiv arXiv 2023

  16. [16]

    S. Lin, J. Hilton, O. Evans, Truthfulqa: Measuring how models mimic human falsehoods, Proceed- ings of the Annual Meeting of the Association for Computational Linguistics 1 (2021) 3214–3252. URL: https://arxiv.org/pdf/2109.07958. doi:10.18653/v1/2022.acl-long.229

    Pith/arXiv arXiv 2021

  17. [17]

    J. Li, X. Cheng, W. X. Zhao, J.-Y. Nie, J.-R. Wen, Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023. URL: http://arxiv.org/abs/2305.11747. doi:10.48550/ arXiv.2305.11747, arXiv:2305.11747 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.11747 2023

  18. [18]

    Exploring the Relationship between LLM Hallucinations and Prompt Linguistic Nuances: Readability, Formality, and Concreteness

    V. Rawte, P. Priya, S. M. Towhidul, I. Tonmoy, S. M. M. Zaman, A. Sheth, A. Das, Exploring the relationship between llm hallucinations and prompt linguistic nuances: Readability, formality, and concreteness (2023). URL: https://arxiv.org/pdf/2309.11064

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    Investigating and Addressing Hallucinations of LLMs in Tasks Involving Negation

    N. Varshney, S. Raj, V. Mishra, A. Chatterjee, R. Sarkar, A. Saeidi, C. Baral, Investigating and addressing hallucinations of llms in tasks involving negation (2024). URL: https://arxiv.org/pdf/ 2406.05494

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Strong hallucinations from negation and how to fix them

    N. Asher, S. Bhar, Strong hallucinations from negation and how to fix them (2024). URL: https: //arxiv.org/pdf/2402.10543

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

    Joshi, S

    A. Joshi, S. Saha, D. Shukla, S. Vema, H. Jhamtani, M. Gaur, A. Modi, Towards robust evaluation of unlearning in llms via data transformations, in: Findings of the Association for Computational Linguistics: EMNLP 2024, 2024, pp. 12100–12119

    work page 2024

  22. [22]

    G. Team, M. Riviere, S. Pathak, P. G. Sessa, C. Hardin, S. Bhupatiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ramé, J. Ferret, P. Liu, P. Tafti, A. Friesen, M. Casbon, S. Ramos, R. Kumar, C. L. Lan, S. Jerome, A. Tsitsulin, N. Vieillard, P. Stanczyk, S. Girgin, N. Momchev, M. Hoffman, S. Thakoor, J.-B. Grill, B. Neyshabur, O. Bachem, A. Walton, A. Sev...

    Pith/arXiv arXiv 2024

  23. [23]

    G. Team, A. Kamath, J. Ferret, S. Pathak, N. Vieillard, R. Merhej, S. Perrin, T. Matejovicova, A. Ramé, M. Rivière, L. Rouillard, T. Mesnard, G. Cideron, J. bastien Grill, S. Ramos, E. Yvinec, M. Casbon, E. Pot, I. Penchev, G. Liu, F. Visin, K. Kenealy, L. Beyer, X. Zhai, A. Tsitsulin, R. Busa- Fekete, A. Feng, N. Sachdeva, B. Coleman, Y. Gao, B. Mustafa,...

    Pith/arXiv arXiv 2025

  24. [24]

    Z. Wu, A. Arora, A. Geiger, Z. Wang, J. Huang, D. Jurafsky, C. D. Manning, C. Potts, Axbench: Steering llms? even simple baselines outperform sparse autoencoders (2025). URL: https://arxiv. org/pdf/2501.17148. A. Online Resources The source code and data related to this work are available at: •GitHub

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.