Pith. sign in

REVIEW 2 major objections 5 minor 211 references

LLM judge bias lives as a low-dimensional direction in the model’s hidden state, and steering along it both creates and cancels unfair scores.

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 · grok-4.5

2026-07-14 02:32 UTC pith:ULXVDONQ

load-bearing objection Solid three-part MI account of LLM-as-judge bias (geometry, bidirectional steering, cross-domain prediction) that earns referee time; main residual is the surface-cue/score-shift operationalization of bias, which the authors partially close with human TOST and out-of-substrate tests. the 2 major comments →

arxiv 2607.11871 v1 pith:ULXVDONQ submitted 2026-07-13 cs.LG cs.AIcs.CL

Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias

Zixiang Xu , Sixian Li , Huaxing Liu , Xiang Wang , Shuai Li , Zirui Song , Xiuying Chen This is my paper
classification cs.LG cs.AIcs.CL
keywords LLM-as-judgescoring biasmechanistic interpretabilityactivation steeringbias subspacerepresentation engineeringoutcome prediction
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.

When large language models are used as automatic judges of answers, their scores move with surface cues that have nothing to do with answer quality—prestige labels, peer consensus notes, length, tone, identity claims. Prior work mostly treated that as black-box input-output noise and tried to fix it with better prompts. This paper argues the same bias has a clear internal geometry. Clean judging inputs sit in a tight activation manifold; biased inputs are displaced along a low-dimensional, type-specific subspace that becomes sharper in deeper layers and is recovered by several independent estimators. Steering the hidden state along that subspace bidirectionally controls scores: adding the direction makes a clean answer look biased, subtracting it restores fair scoring on a biased answer, while matched-norm random directions barely move the score. The same direction features also let a simple linear predictor flag judge failures on three benchmarks never seen during training, beating text-only detectors. If the account holds, bias is no longer only a catalog of prompt tricks but a representation object that can be measured, steered, and predicted.

Core claim

LLM-as-judge scoring bias admits a representation-level account: baseline activations form a tight manifold while biased inputs are displaced along a low-dimensional, type-specific subspace that sharpens with depth; that same subspace is an interventional handle that bidirectionally controls scores and supplies features that predict judge degradation on held-out domains.

What carries the argument

The bias direction (or low-dimensional bias subspace) recovered from effective bias samples via directional-change and discriminative-boundary estimators; unit-normalized and used both for activation steering (add or subtract at mid-to-late layers) and for linear outcome prediction.

Load-bearing premise

The paper treats score shifts caused by carefully constructed surface-cue edits as bias, and fits the direction only on the strong-shift subset of those edits; if many of those cues are rationally quality-relevant or the strong-shift tail is unrepresentative, the geometry may not underwrite the broader fairness claim.

What would settle it

Show that matched-norm random directions or swapped bias-type directions produce score shifts as large as the recovered bias direction under the same validity and rank-preservation constraints, or that a linear projection onto those features fails to beat a strong text baseline on truly held-out domains.

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

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

2 major / 5 minor

Summary. The paper argues that LLM-as-judge scoring bias is a representation-level phenomenon: baseline activations form a tight manifold while biased inputs are displaced along a low-dimensional, type-specific subspace that sharpens with depth. Across seven judges, seven bias types, and nine benchmarks, the authors (i) document strong behavioral asymmetry (negative surface cues penalize more than positive ones reward), (ii) recover the bias subspace with directional and discriminative estimators that agree within family and partially across architectures, (iii) show bidirectional activation steering along that subspace (attack on clean inputs, defense on biased ones) far above matched-norm random and bias-type-swap controls, and (iv) train a linear projection onto the same features that predicts score degradation on three entirely held-out benchmarks (AUC ~0.82 vs ~0.63 text baseline). The contribution is a unified geometric–causal–operational account rather than a new mitigation recipe.

Significance. If the three-part claim holds, the paper supplies a mechanistic account of a widely used evaluation primitive that currently sits inside RLHF and benchmark pipelines. The combination of multi-estimator geometry, bidirectional causal control with random and swap ablations, and a simple transferable linear predictor is stronger than typical input–output bias catalogs and is operationally useful for white-box judges. Strengths include the multi-judge behavioral replication, the TOST human equivalence checks on both prose-rewrite and bit-identical-body perturbations, the matched-budget text-attack comparison, 5-fold CV defense retaining ≥80% of in-sample W1 reduction, and cross-domain prediction sensitivity analysis. These make the result a solid contribution to mechanistic interpretability of evaluators and to fairness auditing of LLM judges.

major comments (2)
  1. [§3.1–3.3, App. F, C/C.1] Sections 3.1–3.3 and Appendix F: direction estimation is restricted to the effective-bias / biased-core case-control subset (score shift ≥ δ_s = 2 and Mahalanobis 90th percentile). The paper correctly treats null-shift samples as null observations and evaluates causal/predictive claims outside the fitting substrate, but the fairness claim still depends on the premise that the surface cues are not rationally quality-relevant. Appendix C/C.1 TOST tests and the matched-budget text comparison (J.3) substantially mitigate this; the manuscript should state more explicitly in the main text (not only the appendix) that the geometry is conditioned on the established surface-cue operationalization of the LLM-as-judge literature, and report a brief sensitivity of recovered directions to δ_s ∈ {1,2,3} so readers can see how much the subspace depends on the strong-shift tail.
  2. [§4.4, App. J.1–J.2] Section 4.4 and Appendices J.1–J.2: bidirectional steering is presented as interventional sufficiency, not unique natural pathway—this scoping is appropriate. The random-direction control is strong (order-of-magnitude gap); the bias-type-swap control correctly shows an intermediate effect (shared subspace + type-specific component). For the central causal claim, the paper should either (a) add a short path-patching / mediation sketch on one bias type and one layer, or (b) keep the current scoping but move the “we do not claim unique pathway” language into the main-text causal paragraph rather than only the discussion, so readers do not over-read the attack/defense results as full causal identification.
minor comments (5)
  1. [Table 2, §4.2] Table 2 vs Table 4 / Figure 3: the positive-aggregate convention (five score-inflating types vs all seven) is explained in Appendix I.1 but should be flagged once in the main-text caption of Table 2 to avoid confusion with the full-pool +0.07 figure mentioned in §4.2.
  2. [Fig. 1–2] Figure 1 MDS and Figure 2 Δh MDS: axis scales differ across panels; a shared color legend and a note that MDS is used for visualization only (not for the estimators) would help non-MI readers.
  3. [App. H, I.3] Appendix H limitations: white-box coverage is limited to three mid-scale judges; the cross-architecture cosine band [0.47, 0.62] is useful—consider promoting one sentence of that partial-transfer result into the main-text geometry section.
  4. [App. G] Reproducibility: seed, split, and ~1400 A100-hour budget are stated; ensure the promised code release includes the exact α-search hyperparameters (Algorithm 1) and the nested question-ID split files so the cross-domain AUC can be regenerated.
  5. [§3.2] Minor notation: ℳ_base is introduced as a manifold but used as an empirical cluster; a one-line clarification that it is the empirical support of H_base^(l) would avoid geometric overclaim.

Circularity Check

0 steps flagged

No significant circularity: geometry, bidirectional steering, and cross-domain prediction are independently validated outside the case-control fitting substrate used to estimate directions.

full rationale

The paper's three claims (activation geometry of bias, bidirectional causal control via steering, and operational prediction of score degradation) are empirical findings from controlled experiments, not derivations that reduce to their inputs by construction. Bias directions are estimated on an effective-bias / biased-core case-control subset defined by observed score shifts (δ_s=2), which is an acknowledged selection of the phenomenon rather than a circular definition of the target quantity; causal attack/defense, random-direction and bias-type-swap controls, 5-fold CV defense, and the linear-projection outcome predictor are all evaluated on inputs outside that substrate (held-out folds, three entirely unseen benchmarks). No equation equates a 'prediction' to a fitted constant; no uniqueness theorem or ansatz is imported via self-citation as a load-bearing premise; self-citations (e.g., to SocialMaze or related representation work by overlapping authors) supply benchmarks or background and are not required for the central results. The operationalization of bias via surface cues is a construct-validity premise (already partially closed by human TOST equivalence tests), not circularity. The work is self-contained against its external benchmarks and controls.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on the linear-representation hypothesis, a surface-cue operationalization of bias taken from prior judge-bias literature, and several numerical thresholds that define the fitting substrate and feasibility set. The 'baseline manifold' and 'bias direction/subspace' are empirical constructs recovered from data rather than postulated particles; their independent handle is the bidirectional steering and cross-domain prediction. Free parameters are mostly thresholds and search bounds that are stated and partially ablated.

free parameters (5)
  • effective-bias score-shift threshold δ_s = 2
    Paired instances enter the direction-fitting substrate only if |Δs| ≥ δ_s integer points in the expected direction; set to 2 throughout.
  • outcome-prediction degradation threshold δ_o = 1
    Positive class for the operational predictor is score drop ≥ δ_o; set more permissively than δ_s.
  • biased-core Mahalanobis percentile = 90th percentile
    𝒟_far retains only the subset of effective negative samples above the 90th percentile of Mahalanobis distance from the baseline cluster.
  • steering strength α* and feasibility floors = per (layer, vector, bias type); V floor 0.93
    α* maximizes Wasserstein shift subject to validity V(α) ≥ 0.93 and Spearman ρ_S(α) ≥ ρ_text_S; localized by two-stage search (~100 forward passes).
  • intervention layer l = dev-set max-separability layer
    Chosen as the layer maximizing biased-vs-baseline separability on the development set (typically mid-to-late, 15–31).
axioms (4)
  • domain assumption Linear representation hypothesis: high-level behaviors (here, typed scoring bias) concentrate along low-dimensional directions or subspaces in residual-stream activations.
    Invoked throughout Sections 3.2–3.4 and 4.3–4.4; standard in the representation-engineering literature the paper builds on.
  • domain assumption Surface-cue operationalization of scoring bias: two inputs that differ only in semantics-irrelevant framing (prestige tags, consensus notes, etc.) and receive systematically different scores constitute measurable bias.
    Section 3.1 and Appendix A.1; adopted from the existing LLM-as-judge bias literature to keep results comparable.
  • ad hoc to paper Case-control framing: null-shift samples are null observations rather than negative examples of the bias phenomenon, so direction estimation may restrict to effective-bias samples.
    Section 3.3; authors defend it as intrinsic to the phenomenon but it shapes the recovered geometry.
  • domain assumption Human raters' quality assessments on template markers and prose rewrites are the ground truth for whether score drops reflect surface bias rather than genuine quality change.
    Appendix C / C.1 TOST equivalence tests; used to support construct validity of the bias labels.
invented entities (2)
  • baseline activation manifold ℳ_base independent evidence
    purpose: Empirical region occupied by unbiased judging activations; biased inputs are defined as geometrically distant from it.
    Recovered from ℋ_base via MDS/PCA; not a new physical entity but a named geometric construct central to the account.
  • type-specific bias direction / low-dimensional bias subspace independent evidence
    purpose: The interventional handle along which biased scoring concentrates; estimated by directional and discriminative families.
    Core object of the paper; independent evidence comes from multi-estimator agreement, bidirectional steering, random/swap controls, and cross-domain prediction.

pith-pipeline@v1.1.0-grok45 · 56494 in / 3780 out tokens · 36383 ms · 2026-07-14T02:32:31.489105+00:00 · methodology

0 comments
read the original abstract

Existing studies of LLM-as-judge scoring bias work predominantly at the input-output level: they perturb inputs, measure score deltas, and propose prompt-level mitigations. We argue that the same biases admit a representation-level account in the judge's hidden state, complementary to the input-output view and operationally useful in ways it does not afford. We report three findings, across seven judges, seven bias types, and nine benchmarks. Geometry: baseline judging inputs occupy a tight activation manifold while biased inputs are displaced along a low-dimensional, type-specific subspace that sharpens with depth and is recovered consistently by three families of estimators. Causal control: steering hidden states along this subspace drives scoring in both directions, forward shifts reproducing biased scoring on clean inputs and reverse shifts restoring baseline scoring on biased ones, while matched-norm random directions produce shifts an order of magnitude smaller. Operational: a simple linear projection onto the same bias-direction features anticipates judge failures on three entirely unseen benchmarks, substantially outperforming text-based alternatives. Reading bias as activation geometry, rather than as input-output noise, unifies geometric structure, causal control, and operational prediction within a single framework. The project page is available at https://xzx34.github.io/unfair-judge/

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

211 extracted references · 3 canonical work pages

  1. [1]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  2. [2]

    2025 , url=

    Qwen2.5 Technical Report , author=. 2025 , url=

    2025

  3. [4]

    2022 , eprint=

    Lin, Stephanie and Hilton, Jacob and Evans, Owain , booktitle=. 2022 , eprint=

    2022

  4. [5]

    2019 , eprint=

    Talmor, Alon and Herzig, Jonathan and Lourie, Nicholas and Berant, Jonathan , booktitle=. 2019 , eprint=

    2019

  5. [6]

    and Lu, Xinghua , booktitle=

    Jin, Qiao and Dhingra, Bhuwan and Liu, Zhengping and Cohen, William W. and Lu, Xinghua , booktitle=. 2019 , eprint=

    2019

  6. [7]

    , booktitle=

    Rein, David and Hou, Betty Li and Stickland, Asa Cooper and Petty, Jackson and Pang, Richard Yuanzhe and Dirani, Julien and Michael, Julian and Bowman, Samuel R. , booktitle=. 2024 , eprint=

    2024

  7. [8]

    2018 , eprint =

    Clark, Peter and Cowhey, Isaac and Etzioni, Oren and Khot, Tushar and Sabharwal, Ashish and Schoenick, Carissa and Tafjord, Oyvind , title =. 2018 , eprint =

    2018

  8. [10]

    2024 , howpublished =

    GPT-4o mini: Advancing Cost-Efficient Intelligence , author =. 2024 , howpublished =

    2024

  9. [11]

    2024 , eprint=

    GPT-4 Technical Report , author=. 2024 , eprint=

    2024

  10. [12]

    2024 , eprint=

    GPT-4o System Card , author=. 2024 , eprint=

    2024

  11. [13]

    2025 , eprint=

    DeepSeek-V3 Technical Report , author=. 2025 , eprint=

    2025

  12. [14]

    Gemma 3 , url=

    Gemma Team , year=. Gemma 3 , url=

  13. [15]

    2024 , month =

    Phi-4 Technical Report , author =. 2024 , month =

    2024

  14. [16]

    2025 , howpublished =

    GPT-4.1 , author =. 2025 , howpublished =

    2025

  15. [17]

    , booktitle=

    Parrish, Alicia and Chen, Angelica and Nangia, Nikita and Padmakumar, Vishakh and Phang, Jason and Thompson, Jana and Htut, Phu Mon and Bowman, Samuel R. , booktitle=. 2022 , eprint=

    2022

  16. [18]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Measuring Massive Multitask Language Understanding , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  17. [19]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle=. Judging. 2023 , eprint=

    2023

  18. [20]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    Large Language Models are not Fair Evaluators , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , year=. 2305.17926 , archivePrefix=

    Pith/arXiv arXiv

  19. [21]

    Findings of the Association for Computational Linguistics: ACL 2024 , year=

    Benchmarking Cognitive Biases in Large Language Models as Evaluators , author=. Findings of the Association for Computational Linguistics: ACL 2024 , year=. 2309.17012 , archivePrefix=

    Pith/arXiv arXiv 2024

  20. [22]

    International Conference on Learning Representations (ICLR) , year=

    Evaluating Large Language Models at Evaluating Instruction Following , author=. International Conference on Learning Representations (ICLR) , year=. 2310.07641 , archivePrefix=

    Pith/arXiv arXiv

  21. [23]

    and Zhang, Xiangliang , booktitle=

    Ye, Jiayi and Wang, Yanbo and Huang, Yue and Chen, Dongping and Zhang, Qihui and Moniz, Nuno and Gao, Tian and Geyer, Werner and Huang, Chao and Chen, Pin-Yu and Chawla, Nitesh V. and Zhang, Xiangliang , booktitle=. Justice or Prejudice? Quantifying Biases in. 2025 , eprint=

    2025

  22. [24]

    Humans or

    Chen, Guiming Hardy and Chen, Shunian and Liu, Ziche and Jiang, Feng and Wang, Benyou , booktitle=. Humans or. 2024 , eprint=

    2024

  23. [25]

    Split and Merge: Aligning Position Biases in

    Li, Zongjie and Wang, Chaozheng and Ma, Pingchuan and Wu, Daoyuan and Wang, Shuai and Gao, Cuiyun and Liu, Yang , booktitle=. Split and Merge: Aligning Position Biases in. 2024 , eprint=

    2024

  24. [26]

    and Feng, Shi , booktitle=

    Panickssery, Arjun and Bowman, Samuel R. and Feng, Shi , booktitle=. 2024 , eprint=

    2024

  25. [27]

    Judging the Judges: Evaluating Alignment and Vulnerabilities in

    Thakur, Aman Singh and Choudhary, Kartik and Ramayapally, Venkat Srinik and Vaidyanathan, Sankaran and Hupkes, Dieuwke , booktitle=. Judging the Judges: Evaluating Alignment and Vulnerabilities in. 2025 , eprint=

    2025

  26. [28]

    A Survey on

    Gu, Jiawei and Jiang, Xuhui and Shi, Zhichao and Tan, Hexiang and Zhai, Xuehao and Xu, Chengjin and Li, Wei and Shen, Yinghan and Ma, Shengjie and Liu, Honghao and Wang, Yuanzhuo and Guo, Jian , year=. A Survey on. 2411.15594 , archivePrefix=

    Pith/arXiv arXiv

  27. [29]

    2412.05579 , archivePrefix=

    Li, Haitao and Dong, Qian and Chen, Junjie and Su, Huixue and Zhou, Yujia and Ai, Qingyao and Ye, Ziyi and Liu, Yiqun , year=. 2412.05579 , archivePrefix=

    Pith/arXiv arXiv

  28. [30]

    2023 , eprint=

    Verbosity Bias in Preference Labeling by Large Language Models , author=. 2023 , eprint=

    2023

  29. [31]

    Length-Controlled

    Dubois, Yann and Galambosi, Bal. Length-Controlled. Conference on Language Modeling (COLM) , year=. 2404.04475 , archivePrefix=

    Pith/arXiv arXiv

  30. [32]

    Self-Preference Bias in

    Wataoka, Koki and Takahashi, Tsubasa and Ri, Ryokan , year=. Self-Preference Bias in. 2410.21819 , archivePrefix=

    Pith/arXiv arXiv

  31. [33]

    2024 , eprint=

    Large Language Models are Inconsistent and Biased Evaluators , author=. 2024 , eprint=

    2024

  32. [34]

    Judging the Judges: A Systematic Study of Position Bias in

    Shi, Lin and Ma, Chiyu and Liang, Weicheng and Ma, Xinyi and Shi, Weijian and Vosoughi, Soroush , booktitle=. Judging the Judges: A Systematic Study of Position Bias in. 2025 , eprint=

    2025

  33. [35]

    2023 , eprint=

    Liu, Yang and Iter, Dan and Xu, Yichong and Wang, Shuohang and Xu, Ruochen and Zhu, Chenguang , booktitle=. 2023 , eprint=

    2023

  34. [36]

    2024 , eprint=

    Wang, Yidong and Yu, Zhuohao and Zeng, Zhengran and Yang, Linyi and Wang, Cunxiang and Chen, Hao and Jiang, Chaoya and Xie, Rui and Wang, Jindong and Xie, Xing and Ye, Wei and Zhang, Shikun and Zhang, Yue , booktitle=. 2024 , eprint=

    2024

  35. [37]

    International Conference on Learning Representations (ICLR) , year=

    Prometheus: Inducing Fine-grained Evaluation Capability in Language Models , author=. International Conference on Learning Representations (ICLR) , year=. 2310.08491 , archivePrefix=

    Pith/arXiv arXiv

  36. [38]

    2025 , eprint=

    Zhu, Lianghui and Wang, Xinggang and Wang, Xinlong , booktitle=. 2025 , eprint=

    2025

  37. [39]

    and Cuadron, Alejandro and Wang, Chenguang and Popa, Raluca Ada and Stoica, Ion , booktitle=

    Tan, Sijun and Zhuang, Siyuan and Montgomery, Kyle and Tang, William Y. and Cuadron, Alejandro and Wang, Chenguang and Popa, Raluca Ada and Stoica, Ion , booktitle=. 2025 , eprint=

    2025

  38. [40]

    and Hajishirzi, Hannaneh , year=

    Lambert, Nathan and Pyatkin, Valentina and Morrison, Jacob and Miranda, LJ and Lin, Bill Yuchen and Chandu, Khyathi and Dziri, Nouha and Kumar, Sachin and Zick, Tom and Choi, Yejin and Smith, Noah A. and Hajishirzi, Hannaneh , year=. 2403.13787 , archivePrefix=

    Pith/arXiv arXiv

  39. [41]

    Preference Leakage: A Contamination Problem in

    Li, Dawei and Sun, Renliang and Huang, Yue and Zhong, Ming and Jiang, Bohan and Han, Jiawei and Zhang, Xiangliang and Wang, Wei and Liu, Huan , year=. Preference Leakage: A Contamination Problem in. 2502.01534 , archivePrefix=

    arXiv

  40. [42]

    Replacing Judges with Juries: Evaluating

    Verga, Pat and Hofstatter, Sebastian and Althammer, Sophia and Su, Yixuan and Piktus, Aleksandra and Arkhangorodsky, Arkady and Xu, Minjie and White, Naomi and Lewis, Patrick , year=. Replacing Judges with Juries: Evaluating. 2404.18796 , archivePrefix=

    Pith/arXiv arXiv

  41. [43]

    Beyond the Surface: Enhancing

    Lai, Peng and Zheng, Jianjie and Cheng, Sijie and Chen, Yun and Li, Peng and Liu, Yang and Chen, Guanhua , year=. Beyond the Surface: Enhancing. 2508.03550 , archivePrefix=

    arXiv

  42. [44]

    2023 , eprint=

    Activation Addition: Steering Language Models Without Optimization , author=. 2023 , eprint=

    2023

  43. [45]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=. 2306.03341 , archivePrefix=

    Pith/arXiv arXiv

  44. [46]

    and Wang, Zifan and Mallen, Alex and Basart, Steven and Koyejo, Sanmi and Song, Dawn and Fredrikson, Matt and Kolter, J

    Zou, Andy and Phan, Long and Chen, Sarah and Campbell, James and Guo, Phillip and Ren, Richard and Pan, Alexander and Yin, Xuwang and Mazeika, Mantas and Dombrowski, Ann-Kathrin and Goel, Shashwat and Li, Nathaniel and Byun, Michael J. and Wang, Zifan and Mallen, Alex and Basart, Steven and Koyejo, Sanmi and Song, Dawn and Fredrikson, Matt and Kolter, J. ...

    Pith/arXiv arXiv

  45. [47]

    Steering

    Panickssery, Nina and Gabrieli, Nick and Schulz, Julian and Tong, Meg and Hubinger, Evan and Turner, Alexander Matt , booktitle=. Steering. 2024 , eprint=

    2024

  46. [48]

    Findings of the Association for Computational Linguistics: ACL 2022 , year=

    Extracting Latent Steering Vectors from Pretrained Language Models , author=. Findings of the Association for Computational Linguistics: ACL 2022 , year=. 2205.05124 , archivePrefix=

    Pith/arXiv arXiv 2022

  47. [49]

    International Conference on Learning Representations (ICLR) , year=

    Discovering Latent Knowledge in Language Models Without Supervision , author=. International Conference on Learning Representations (ICLR) , year=. 2212.03827 , archivePrefix=

    Pith/arXiv arXiv

  48. [50]

    Conference on Language Modeling (COLM) , year=

    The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets , author=. Conference on Language Modeling (COLM) , year=. 2310.06824 , archivePrefix=

    Pith/arXiv arXiv

  49. [51]

    International Conference on Machine Learning (ICML) , year=

    The Linear Representation Hypothesis and the Geometry of Large Language Models , author=. International Conference on Machine Learning (ICML) , year=. 2311.03658 , archivePrefix=

    Pith/arXiv arXiv

  50. [52]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Refusal in Language Models Is Mediated by a Single Direction , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=. 2406.11717 , archivePrefix=

    Pith/arXiv arXiv

  51. [53]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , booktitle=. Locating and Editing Factual Associations in. 2022 , eprint=

    2022

  52. [54]

    International Conference on Learning Representations (ICLR) , year=

    Progress Measures for Grokking via Mechanistic Interpretability , author=. International Conference on Learning Representations (ICLR) , year=. 2301.05217 , archivePrefix=

    Pith/arXiv arXiv

  53. [55]

    Proceedings of the 6th BlackboxNLP Workshop , year=

    Emergent Linear Representations in World Models of Self-Supervised Sequence Models , author=. Proceedings of the 6th BlackboxNLP Workshop , year=. 2309.00941 , archivePrefix=

    Pith/arXiv arXiv

  54. [56]

    The Internal State of an

    Azaria, Amos and Mitchell, Tom , booktitle=. The Internal State of an. 2023 , eprint=

    2023

  55. [57]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Inspecting and Editing Knowledge Representations in Language Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=. 2304.00740 , archivePrefix=

    Pith/arXiv arXiv

  56. [58]

    and McDougall, Callum and MacDiarmid, Monte and Tamkin, Alex and Durmus, Esin and Hume, Tristan and Mosconi, Francesco and Freeman, C

    Templeton, Adly and Conerly, Tom and Marcus, Jonathan and Lindsey, Jack and Bricken, Trenton and Chen, Brian and Pearce, Adam and Citro, Craig and Ameisen, Emmanuel and Jones, Andy and Cunningham, Hoagy and Turner, Nicholas L. and McDougall, Callum and MacDiarmid, Monte and Tamkin, Alex and Durmus, Esin and Hume, Tristan and Mosconi, Francesco and Freeman...

  57. [59]

    2023 , howpublished=

    Towards Monosemanticity: Decomposing Language Models With Dictionary Learning , author=. 2023 , howpublished=

    2023

  58. [60]

    and Espinosa-Anke, Luis , year=

    Siddique, Zara and Turner, Liam D. and Espinosa-Anke, Luis , year=. Shifting Perspectives: Steering Vector Ensembles for Robust Bias Mitigation in. 2503.05371 , archivePrefix=

    arXiv

  59. [61]

    2025 , eprint=

    Representation Engineering for Large-Language Models: Survey and Research Challenges , author=. 2025 , eprint=

    2025

  60. [62]

    Activation Steering for Bias Mitigation: An Interpretable Approach to Safer

    Dubey, Shivam , year=. Activation Steering for Bias Mitigation: An Interpretable Approach to Safer. 2508.09019 , archivePrefix=

    Pith/arXiv arXiv

  61. [63]

    International Conference on Machine Learning (ICML) , year=

    Discovering Bias in Latent Space: An Unsupervised Debiasing Approach , author=. International Conference on Machine Learning (ICML) , year=. 2406.03631 , archivePrefix=

    Pith/arXiv arXiv

  62. [64]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year=. 2004.07667 , archivePrefix=

    Pith/arXiv arXiv 2004

  63. [65]

    International Conference on Machine Learning (ICML) , year=

    Linear Adversarial Concept Erasure , author=. International Conference on Machine Learning (ICML) , year=. 2201.12091 , archivePrefix=

    Pith/arXiv arXiv

  64. [66]

    2023 , eprint=

    Belrose, Nora and Schneider-Joseph, David and Ravfogel, Shauli and Cotterell, Ryan and Raff, Edward and Biderman, Stella , booktitle=. 2023 , eprint=

    2023

  65. [67]

    International Conference on Learning Representations (ICLR) , year=

    Eliminating Position Bias of Language Models: A Mechanistic Approach , author=. International Conference on Learning Representations (ICLR) , year=. 2407.01100 , archivePrefix=

    Pith/arXiv arXiv

  66. [68]

    International Conference on Learning Representations (ICLR) , year=

    Programming Refusal with Conditional Activation Steering , author=. International Conference on Learning Representations (ICLR) , year=. 2409.05907 , archivePrefix=

    Pith/arXiv arXiv

  67. [69]

    2025 , eprint=

    Li, Haitao and Chen, Junjie and Ai, Qingyao and Chu, Zhumin and Zhou, Yujia and Dong, Qian and Liu, Yiqun , booktitle=. 2025 , eprint=

    2025

  68. [70]

    Adaptive Distraction: Probing

    Wang, Yanbo and Xu, Zixiang and Huang, Yue and Gao, Chujie and Wu, Siyuan and Ye, Jiayi and Chen, Pin-Yu and Chen, Xiuying and Zhang, Xiangliang , booktitle =. Adaptive Distraction: Probing

  69. [71]

    Wang, Yanbo and Xu, Zixiang and Huang, Yue and Wang, Xiangqi and Song, Zirui and Gao, Lang and Wang, Chenxi and Tang, Robert and Zhao, Yue and Cohan, Arman and Zhang, Xiangliang and Chen, Xiuying , booktitle =

  70. [74]

    Xu, Zixiang and Wang, Yanbo and Huang, Yue and Wang, Chenxi and Gao, Lang and Song, Zirui and Chen, Zhaorun and Zhang, Xiangliang and Chen, Xiuying , year=

  71. [78]

    Proceedings of the 43rd International Conference on Machine Learning , volume=

    Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments , author=. Proceedings of the 43rd International Conference on Machine Learning , volume=. 2026 , eprint=

    2026

  72. [81]

    Advances in Neural Information Processing Systems , pages=

    Hu, Yikuan and Zhu, Jifeng and Tang, Lanrui and Huang, Chen , editor=. Advances in Neural Information Processing Systems , pages=. 2025 , url=

    2025

  73. [85]

    Preprints , doi=

    The Evaluation Bottleneck of Vision-Language-Action Models: An Evaluation-Centric Survey , author=. Preprints , doi=

  74. [86]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Shaping the Safety Boundaries: Understanding and Defending Against Jailbreaks in Large Language Models , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

  75. [87]

    2025 , eprint=

    Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMs , author=. 2025 , eprint=

    2025

  76. [88]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    When Personalization Tricks Detectors: The Feature-Inversion Trap in Machine-Generated Text Detection , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

  77. [89]

    2026 , eprint=

    The Cylindrical Representation Hypothesis for Language Model Steering , author=. 2026 , eprint=

    2026

  78. [90]

    Decoding Echo Chambers:

    Wang, Chenxi and Liu, Zongfang and Yang, Dequan and Chen, Xiuying , booktitle =. Decoding Echo Chambers:. 2025 , address =

    2025

  79. [91]

    Word Form Matters:

    Wang, Chenxi and Gu, Tianle and Wei, Zhongyu and Gao, Lang and Song, Zirui and Chen, Xiuying , booktitle =. Word Form Matters:. 2025 , address =

    2025

  80. [92]

    2025 , eprint=

    Under the Shadow of Babel: How Language Shapes Reasoning in LLMs , author=. 2025 , eprint=

    2025

Showing first 80 references.