Pith. sign in

REVIEW 4 major objections 4 minor 61 references

Editing an AI's reasoning graph lifts accuracy by up to 24 points

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 ↗

An interactive chain-of-thought editor with prune and graft interventions reports large accuracy gains, but without error bars, code, or a comparison to existing editing tools.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A promising human-in-the-loop CoT editing system whose central DAG claim is undercut by its own formal definition of the graph as a linear chain. the 4 major comments →

arxiv 2509.01412 v3 pith:UQAUXDPB submitted 2025-09-01 cs.CL

Vis-CoT: A Human-in-the-Loop Framework for Interactive Visualization and Intervention in LLM Chain-of-Thought Reasoning

Kaviraj Pather , Elena Hadjigeorgiou , Arben Krasniqi , Claire Schmit , Irina Rusu , Marc Pons , Kabir Khan This is my paper
classification cs.CL
keywords chain-of-thought reasoninghuman-in-the-loopinteractive visualizationexplainable AIlarge language modelsreasoning graphLLM interventiontrust in AI
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 reading

Vis-CoT is a human-in-the-loop framework that turns the linear text of an LLM's chain-of-thought reasoning into an interactive graph. Users can see each reasoning step, mark bad steps, prune entire faulty sub-paths, and graft in their own premises; the edited graph is fed back to the model to continue reasoning. The paper reports that this collaborative loop lifts final-answer accuracy by up to 24 percentage points over non-interactive baselines on GSM8K, StrategyQA, and a custom planning task, and that users report much higher usability and trust. The claim is that many LLM reasoning failures are localized and correctable, so targeted human oversight can be a practical path to more reliable and controllable AI without retraining the model.

Core claim

The central claim is that the opacity of chain-of-thought reasoning is not a fixed limit: if the reasoning trace is represented as a directed acyclic graph G = (V, E), with nodes as steps and edges as dependencies, a user can localize a flawed step and apply three operations—flag, prune, and graft—that rewrite the graph. The Model Feedback Module then constructs a prompt from the validated path plus the user's new premise and asks the LLM to continue from that corrected frontier. On GSM8K, Vis-CoT reaches 91.7% accuracy versus 74.8% for standard CoT; on StrategyQA it reaches 94.1% versus 79.5%; on the custom planning task it reaches 92.0% versus 68.0%. The accompanying user study reports Sys

What carries the argument

The load-bearing object is the reasoning graph G = (V, E), a directed acyclic graph in which each node stores a reasoning step's text, a confidence score (average log-probability), a state, and a type, and each directed edge means 'is a direct logical consequence of.' The intervention operations Prune(vi) removes a node and everything reachable from it, while Graft(vp, s_new) adds a user-authored node under a chosen parent. The Model Feedback Module turns the surviving valid path plus the new premise into a continuation prompt, so the LLM resumes from the human-corrected point. This mechanism converts a monolithic text trace into a searchable, editable object, enabling surgical correction in

Load-bearing premise

The load-bearing premise is that the CoT Structuring Module's graph edges are faithful to the model's actual logical dependencies; this parsing accuracy is never evaluated in the paper.

What would settle it

Ask two independent annotators to mark the true dependency edges on 100 chain-of-thought traces, then compare their edge sets with the DAG produced by the CoT Structuring Module; if agreement is low, the accuracy and trust gains cannot be attributed to the visualized structure. A second check: rerun the user study with randomly shuffled edges, and if gains persist, the graph structure is not doing the work.

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

If this is right

  • If most LLM reasoning errors are localized, targeted human edits can rescue a derivation; further accuracy gains may come from better human-AI interfaces rather than larger models.
  • The same interactive loop should transfer to multi-step planning and constraint-heavy tasks, where cascading errors are common and the largest reported gain was observed.
  • Displaying reasoning as a graph makes logical leaps and irrelevant steps visible, so tasks whose traces are currently hidden in dense text should benefit from the representation change.
  • The trust gains suggest that users accept AI outputs they can audit and co-author; high-stakes deployment may need transparency and control alongside raw accuracy.

Where Pith is reading between the lines

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

  • A natural extension is to replace the human with a retrieval step: if grafting works because users inject external facts, then grounding the graph with a knowledge base might deliver similar accuracy gains without a human in the loop.
  • The reported gains come from 24 CS/related participants; whether they generalize to domain experts or novices is untested, and a stratified user study would settle it.
  • Because the parser's faithfulness is unmeasured, the safest interpretation is that the framework demonstrates correction of errors in the parsed graph; real CoT traces with ambiguous dependencies may yield smaller gains.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 4 minor

Summary. The paper introduces Vis-CoT, a human-in-the-loop framework that parses a large language model's chain-of-thought (CoT) trace into a graph, visualizes it, and allows users to flag, prune, or graft reasoning steps. The modified graph is compiled into a new prompt that continues the reasoning, iterating until the user is satisfied. The authors evaluate Vis-CoT on GSM8K, StrategyQA, and a custom multi-step planning dataset, reporting large accuracy gains over non-interactive baselines (up to 24 percentage points) and a user study showing improved SUS and trust scores. The central claim is that converting linear CoT into an interactive DAG enables targeted correction of localized reasoning errors.

Significance. If the DAG-based intervention mechanism were actually implemented and the empirical results robust, this would be a useful contribution to human-in-the-loop reasoning and LLM interpretability. The paper has strengths: it provides a concrete algorithmic description (Algorithm 1), explicit definitions of the three intervention operators, and a worked case study. The idea that users can fix localized errors without full regeneration is intuitively appealing and aligns with current interest in interactive XAI. However, as detailed below, the formalization does not deliver the promised DAG structure, and the empirical evidence is far too thin to support the claimed gains. With substantial revision—particularly a validated graph extraction procedure, a properly reported user study, and release of datasets/code—the underlying idea could be salvageable.

major comments (4)
  1. [Section III-B, Eq. (1)] The formal definition of the reasoning graph is a linear chain, not a DAG with logical dependency edges. Eq. (1) sets E = {(vi, vi+1) | i = 1,...,n−1}, i.e., only consecutive steps. Consequently, Prune (Eq. 4) always removes a suffix of the chain (the reachable subgraph of any node in a directed path is exactly the remaining path), and Graft (Eq. 5) attaches a new node to exactly one parent, with no branching. The paper's repeated claims of pruning 'entire fallacious sub-trees' (Section V-A) and of a 'global graph view' revealing non-linear dependencies (Section V-A) are therefore unsupported by the formalization. This is not merely an evaluation gap; as defined, the system does not implement an interactive DAG. The authors must either extend Eq. (1) to allow arbitrary dependency edges and validate the extraction procedure, or reframe the contribution as step-level editing of a linear tr
  2. [Section IV-A (Custom Planning Dataset)] The custom planning dataset is described only as '100 multi-step planning scenarios' with examples such as conference scheduling, travel itinerary planning, and event resource allocation. There is no description of how the scenarios were generated, their difficulty distribution, the number of gold-standard solutions, or any example. Since the largest reported accuracy gain (24 points over Standard CoT, Table I) comes from this dataset, the result is impossible to interpret or reproduce. The authors should release the dataset or at least provide a full description with several representative examples, and ideally an annotation protocol.
  3. [Section V-A and Section IV-C (Statistical evidence)] All empirical results—accuracy in Table I, efficiency in Table II, and user-experience scores in Table III—are reported as point estimates with no error bars, confidence intervals, or statistical tests. The user study has only 24 participants, and Section IV-B states they were 'randomly assigned' but does not report how many were in each condition, whether the assignment was balanced, or how intervention correctness was verified. The StrategyQA and custom-planning accuracies rely on human evaluators, but no inter-annotator agreement is given. Without these, the claimed gains (e.g., SUS 88.2 vs. 65.5, Trust 4.6 vs. 2.8) could easily be within noise. The authors should report per-condition sample sizes, standard deviations, and appropriate significance tests (e.g., nonparametric paired tests for SUS/trust), and provide the evaluation rubric and annotation agreement for the human-judged tas
  4. [Section III-B (CoT Structuring Module)] The core prerequisite of the framework—parsing raw CoT text into a faithful graph G=(V,E)—is never evaluated. The paper gives no procedure for extracting edges beyond consecutive steps, and no experiment compares the extracted graph to a human-annotated gold graph (e.g., in terms of edge precision/recall or downstream accuracy). If the parser mis-segments steps or produces edges that do not reflect actual logical dependencies, then the entire intervention mechanism operates on an unreliable representation and the measured accuracy/trust gains could be artifacts of the specific (possibly chain-like) outputs of Llama-2-70B-Chat on these benchmarks. The authors should add a faithfulness evaluation and, at minimum, an ablation comparing the full graph mechanism against a simple linear-trace editor to isolate the benefit of the graph structure.
minor comments (4)
  1. [Author affiliations] In the author block, Kabir Khan's affiliation lists 'San Francisco State University, San Francisco, CA 94132, India'; the location should be 'USA'.
  2. [References] Reference [13] appears garbled: the author list contains 'Izhak Sha, Thomas Pang, Izhak...' and 'Zheru Ranalde', which are not valid author names. Several other references (e.g., [20], [22]) appear to have been generated with inconsistent metadata. The reference for the trust scale, [60], is actually a survey on CoT interpretability, not the source of the adapted scale; please cite the actual scale.
  3. [Figures] Figures 1–4 are referenced in the text but are not included in the provided manuscript. The case study (Section V-C) would be much stronger if it included the before/after graph visualization. Ensure all figures are embedded and legible, and label the specific components of the interface described in Sections III-C and III-D.
  4. [Section V-A, Table II] The average number of interventions is reported as 2.1 with no measure of spread. Given the small user-study sample, report the distribution (e.g., standard deviation, range, or interquartile range) to give a sense of variability.

Circularity Check

2 steps flagged

The 'interactive reasoning graph' is defined by Eq. (1) as a linear chain, so the claimed DAG-based visualization and pruning/grafting mechanism reduce by construction to editing the original CoT sequence.

specific steps
  1. self definitional [Section III-B, Eq. (1)]
    "Each node vi ∈ V corresponds to the i-th step in the reasoning chain. An edge ei,j ∈ E from node vi to vj indicates that step j is a direct logical consequence of step i. For a standard linear CoT, this graph is a simple chain. ... G = (V, E) s.t. V = {v1, v2, . . . , vn}, E = {(vi, vi+1) | i = 1, . . . , n−1}"

    The 'reasoning graph' is defined with edges only between consecutive steps, so G is identical to the linear CoT input, just relabeled. No procedure is given for extracting non-consecutive dependency edges. Therefore the claimed conversion of linear CoT into a DAG with logical-dependency edges is not an independent derivation; the graph's structure is the input by construction. Any subsequent claim about a 'global graph view' revealing non-linear dependencies or circular leaps is unsupported by Eq. (1).

  2. self definitional [Section III-D, Eqs. (4)-(5)]
    "Prune(vi) → G′ = (V \ Vsub(vi), E \ Esub(vi)) (4) ... Graft(vp, snew) → V ′ = V ∪ {vnew}, E ′ = E ∪ {(vp, vnew)} (5)"

    Because Eq. (1) makes G a chain, Vsub(vi) is necessarily the suffix {vi,...,vn} and there are no branches or sub-trees. Pruning is suffix deletion; grafting adds a single node to one parent, which can only create a branch if the user first prunes the chain. The paper's qualitative claims of pruning 'entire fallacious sub-trees' and of a branching, editable reasoning structure are thus not implementations of a general DAG intervention mechanism; they reduce by construction to editing a linear sequence. The framework's formal mechanism does not realize the DAG-based intervention it claims.

full rationale

The paper's empirical accuracy comparisons on GSM8K and StrategyQA are externally measured, not derived from the framework's equations, and there is no load-bearing self-citation chain: the named authors do not overlap with the cited prior work, and the benchmarks are standard. However, the paper's formal 'reasoning graph' is defined in Eq. (1) with edge set {(vi, vi+1)}, which is exactly the linear CoT sequence renamed as a graph. The claimed structuring step—converting linear CoT into an interactive DAG with dependency edges and sub-trees—is therefore not an extraction or prediction; it is a relabeling by construction. The intervention operations (Prune and Graft) then operate on that chain, making 'pruning sub-trees' and 'global graph view' claims unsupported by the formal definitions. This is a partial circularity: the central representational contribution reduces to its input, while the human-in-the-loop accuracy gains remain an empirical result about step-level editing, not about a DAG-based intervention mechanism. Score 6 reflects that one core structural claim is definitionally identical to its input, even though the benchmark results and user-study findings are not circular in themselves.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The framework relies on several domain assumptions that are not independently validated in the paper: faithful CoT parsing into a DAG, user ability to supply correct corrections, and stable LLM continuation from an edited graph. There are no fitted numeric parameters in the modeling, and no new physical or conceptual entities are postulated.

axioms (4)
  • domain assumption CoT text can be parsed faithfully into a DAG where each edge represents a genuine logical dependency.
    Section III-B Eq. (1) defines G=(V,E) with E={(vi,vi+1)} for basic CoT, but parser accuracy and edge faithfulness are never evaluated; all pruning and grafting operate on this graph.
  • domain assumption Users can identify incorrect reasoning steps and supply correct premises.
    Section V-D and the Limitations section state effectiveness depends on the user's ability to identify errors, but no measure of user expertise or intervention correctness is reported.
  • domain assumption The LLM will continue coherently from a user-grafted node using the concatenated validated path and new premise.
    Section III-E Eq. (6) builds Cprompt and asks the model to continue from the edited graph; no studies quantify continuation quality or failure modes.
  • domain assumption The Standard CoT baseline is a representative non-interactive practice condition.
    Section IV-D defines the baseline, but no seed protocol, prompt details, or multiple runs are reported, so the comparison could depend on unstated implementation choices.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Vis-CoT: A Human-in-the-Loop Framework for Interactive Visualization and Intervention in LLM Chain-of-Thought Reasoning." pith.science (2026). https://pith.science/paper/UQAUXDPB

@misc{pith2026250901412,
  author       = {Pith},
  title        = {Pith review of: Vis-CoT: A Human-in-the-Loop Framework for Interactive Visualization and Intervention in LLM Chain-of-Thought Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQAUXDPB}},
  note         = {Machine review of arXiv:2509.01412}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) show strong reasoning via chain-of-thought (CoT) prompting, but the process is opaque, which makes verification, debugging, and control difficult in high-stakes settings. We present Vis-CoT, a human-in-the-loop framework that converts linear CoT text into an interactive reasoning graph. Users can visualize the logical flow, identify flawed steps, and intervene by pruning incorrect paths and grafting new, user-defined premises. This shifts interaction from passive observation to active collaboration, steering models toward more accurate and trustworthy conclusions. Across GSM8K and StrategyQA, Vis-CoT improves final-answer accuracy by up to 24 percentage points over non-interactive baselines. A user study also shows large gains in perceived usability and trust. Vis-CoT points to a practical path for more reliable, understandable, and collaborative reasoning by combining LLMs with targeted human oversight.

Figures

Figures reproduced from arXiv: 2509.01412 by Arben Krasniqi, Claire Schmit, Elena Hadjigeorgiou, Irina Rusu, Kabir Khan, Kaviraj Pather, Marc Pons.

Figure 1
Figure 1. Figure 1: Overall architecture of the Vis-CoT framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Three intervention operations: (a) Flagging, (b) Prun [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Efficiency comparison before/after intervention (lower [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Final Answer Accuracy on GSM8K, StrategyQA, and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case study visualization (Before vs. After intervention). [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] 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

61 extracted references · 46 canonical work pages · 17 internal anchors

  1. [1]

    Emergent abilities of large language models

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebas- tian Borgeaud, Dani Yogatama, William Fedus, Aakanksha Chowdhery, Sharan Narang, et al. Emergent abilities of large language models. Transactions on Machine Learning Research , 2022

    work page 2022

  2. [2]

    Context-aware graph inference with knowledge distillation for visual dialog

    Dan Guo, Hui Wang, and Meng Wang. Context-aware graph inference with knowledge distillation for visual dialog. IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(10):6056–6073, 2021

    work page 2021

  3. [3]

    Active velocity processes with suprathermal stationary distributions and long-time tails

    Daniel Keysers, Nathanael Sch ¨arli, Jakub Raiman, Ankur Bapna, Quoc V Le, and Orhan Firat. Measuring and narrowing the compo- sitionality gap in language models. arXiv preprint arXiv:1912.10780 , 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  4. [4]

    Federated learning with experience- driven model migration in heterogeneous edge networks

    Jianchun Liu, Shilong Wang, Hongli Xu, Yang Xu, Yunming Liao, Jinyang Huang, and He Huang. Federated learning with experience- driven model migration in heterogeneous edge networks. IEEE/ACM Transactions on Networking , 32(4):3468–3484, 2024

    work page 2024

  5. [5]

    Gamut: A design probe for understanding developer challenges in building intelligent interactive systems

    Fred Hohman, Andrew Head, Mary Beth Kery, and for the Projects. Gamut: A design probe for understanding developer challenges in building intelligent interactive systems. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pages 1–13, 2019

    work page 2019

  6. [6]

    NISF: Neural Implicit Segmentation Functions

    Zhaowei Zhang, Zhaofeng Zhang, Zhaoxuan Chen, Zhaofeng Feng, Xiaoya Li, Kanyuan Zhou, Zhipeng Li, Jing Wang, and Yiquan Wang. Evaluating and improving the chain-of-thought reasoning in large lan- guage models. arXiv preprint arXiv:2309.08643 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Multi-objective convex quantization for efficient model compression

    Chunxiao Fan, Dan Guo, Ziqi Wang, and Meng Wang. Multi-objective convex quantization for efficient model compression. IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    work page 2024

  8. [8]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems , 35:24824–24837, 2022

    work page 2022

  9. [9]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in Neural Information Processing Systems , 35:22199–22213, 2022

    work page 2022

  10. [10]

    Resup: Reliable label noise suppression for facial expression recognition

    Xiang Zhang, Yan Lu, Huan Yan, Jinyang Huang, Yu Gu, Yusheng Ji, Zhi Liu, and Bin Liu. Resup: Reliable label noise suppression for facial expression recognition. IEEE Transactions on Affective Computing , pages 1–14, 2025

    work page 2025

  11. [11]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022

    Pith/arXiv arXiv 2022

  12. [12]

    Phaseanti: An anti- interference wifi-based activity recognition system using interference- independent phase component

    Jinyang Huang, Bin Liu, Chenglin Miao, Yan Lu, Qijia Zheng, Yu Wu, Jiancun Liu, Lu Su, and Chang Wen Chen. Phaseanti: An anti- interference wifi-based activity recognition system using interference- independent phase component. IEEE Transactions on Mobile Comput- ing, 22(5):2938–2954, 2023

    work page 2023

  13. [13]

    Tree of thoughts: Deliberate problem solving with large language models

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Sha, Thomas Pang, Jiaming Chen, Zheru Ranalde, Yann Pope, Aohua Chen, Luyu Wang, et al. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 , 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Graph-of-thoughts: Solving elaborate problems with large language models

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Nisa, and Torsten Hoefler. Graph-of-thoughts: Solving elaborate problems with large language models. arXiv preprint arXiv:2308.09687 , 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Measuring faithfulness in chain-of-thought reasoning

    Tamera Lanham, Anna Botev, Jiri Hron, Laurence Aitchison, Adam Clark, J ¨org Schuett, and Mark van der Wilk. Measuring faithfulness in chain-of-thought reasoning. arXiv preprint arXiv:2307.13702 , 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Contrastive positive sample propagation along the audio-visual event line

    Jinxing Zhou, Dan Guo, and Meng Wang. Contrastive positive sample propagation along the audio-visual event line. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(6):7239–7257, 2022

    work page 2022

  17. [17]

    The internal state of an llm knows when its lying

    Amos Azaria and Tom Mitchell. The internal state of an llm knows when its lying. arXiv preprint arXiv:2304.13734 , 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Wiopen: A robust wi-fi-based open-set gesture recognition framework

    Xiang Zhang, Jinyang Huang, Huan Yan, Yuanhao Feng, Peng Zhao, Guohang Zhuang, Zhi Liu, and Bin Liu. Wiopen: A robust wi-fi-based open-set gesture recognition framework. IEEE Transactions on Human- Machine Systems, 55(2):234–245, 2025

    work page 2025

  19. [19]

    Wi- pulmo: Commodity wifi can capture your pulmonary function without mouth clinging

    Peng Zhao, Jinyang Huang, Xiang Zhang, Zhi Liu, Huan Yan, Meng Wang, Guohang Zhuang, Yutong Guo, Xiao Sun, and Meng Li. Wi- pulmo: Commodity wifi can capture your pulmonary function without mouth clinging. IEEE Internet of Things Journal , 12(1):854–868, 2025

    work page 2025

  20. [20]

    Underwater Sound Speed Profile Construction: A Review

    Zhaoxuan Wu, Zhaofeng Feng, He Zhang, Zhaoxuan Chen, Xiaoya Li, Kanyuan Zhou, Zhipeng Li, Jing Wang, and Yiquan Wang. Self- critique and self-correction for large language models. arXiv preprint arXiv:2310.08251, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  21. [21]

    Target-oriented wifi sensing for respiratory healthcare: from indiscriminate perception to in- area sensing

    Meng Wang, Jinyang Huang, Xiang Zhang, Zhi Liu, Meng Li, Peng Zhao, Huan Yan, Xiao Sun, and Mianxiong Dong. Target-oriented wifi sensing for respiratory healthcare: from indiscriminate perception to in- area sensing. IEEE Network, pages 1–1, 2024

    work page 2024

  22. [22]

    Reflexion: an autonomous agent with dynamic memory and self-reflection

    Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with dynamic memory and self-reflection. arXiv preprint arXiv:2303.11366, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Beyond chain-of-thought, effective graph-of-thought reasoning in large language models

    Yao Zhou, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. Beyond chain-of-thought, effective graph-of-thought reasoning in large language models. arXiv preprint arXiv:2305.16582 , 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Phyfinatt: An undetectable attack framework against phy layer fingerprint-based wifi authentication

    Jinyang Huang, Bin Liu, Chenglin Miao, Xiang Zhang, Jianchun Liu, Lu Su, Zhi Liu, and Yu Gu. Phyfinatt: An undetectable attack framework against phy layer fingerprint-based wifi authentication. IEEE Transac- tions on Mobile Computing , 23(7):7753–7770, 2024

    work page 2024

  25. [25]

    The Random-Bond Ising Model and its dual in Hyperbolic Spaces

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Causal tracing for large language models. arXiv preprint arXiv:2210.07227 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  26. [26]

    El-Ad May, Tal Geva, Tom Mitchell, and Jonathan Berant

    Mor Geva, Avi Caciularu, A. El-Ad May, Tal Geva, Tom Mitchell, and Jonathan Berant. Inference-time intervention: Eliciting desired behaviors from llms. arXiv preprint arXiv:2306.03341 , 2023

    Pith/arXiv arXiv 2023

  27. [27]

    Wife: Wifi and vision based unobtrusive emotion recognition via gesture and facial expression

    Yu Gu, Xiang Zhang, Huan Yan, Jingyang Huang, Zhi Liu, Mianxiong Dong, and Fuji Ren. Wife: Wifi and vision based unobtrusive emotion recognition via gesture and facial expression. IEEE Transactions on Affective Computing, 14(4):2567–2581, 2023

    work page 2023

  28. [28]

    Reason: An interactive system for reasoning with and about rationales

    Swarnadeep Wiegreffe, Jack Hessel, Ari Holtzman, Yejin Choi, and Rowan Zellers. Reason: An interactive system for reasoning with and about rationales. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: System Demonstrations , pages 58–69, 2022

    work page 2022

  29. [29]

    Generalize Polyp Segmentation via Inpainting across Diverse Backgrounds and Pseudo-Mask Refinement

    Zhibin Zhang, Shijun Wang, Zichao Wang, Yione Zhang, Yida Wang, Yingcai Wu, and Wei Zhang. Mind’s eye: A visualization and editing tool for chain-of-thought reasoning. arXiv preprint arXiv:2405.12784 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    Learning from heterogeneity: Generalizing dynamic facial expression recognition via distributionally robust opti- mization

    Feng-Qi Cui, Anyang Tong, Jinyang Huang, Jie Zhang, Dan Guo, Zhi Liu, and Meng Wang. Learning from heterogeneity: Generalizing dynamic facial expression recognition via distributionally robust opti- mization. In Proceedings of the 33nd ACM International Conference on Multimedia , MM ’25, New York, NY , USA, 2025. Association for Computing Machinery

    work page 2025

  31. [31]

    Take a step back: Evoking reasoning via abstraction in large language models

    Huaixiu Steven Zheng, Swaroop Swamy, Saurabh Arora, Yiding Gu, Zongyan Liu, Huaxiu Song, Spandan Sahoo, Jure Leskovec, Boris Zaslavskiy, and Anima Lamb. Take a step back: Evoking reasoning via abstraction in large language models. arXiv preprint arXiv:2310.06117, 2023

    Pith/arXiv arXiv 2023

  32. [32]

    Unified static and dynamic network: efficient temporal filtering for video grounding

    Jingjing Hu, Dan Guo, Kun Li, Zhan Si, Xun Yang, Xiaojun Chang, and Meng Wang. Unified static and dynamic network: efficient temporal filtering for video grounding. IEEE Transactions on Pattern Analysis and Machine Intelligence , 2025

    work page 2025

  33. [33]

    Program of thoughts prompting: Disentangling computation from reasoning

    Wenhu Chen, Mohit Bansal, Nate Kushman, and Denny Smile. Program of thoughts prompting: Disentangling computation from reasoning. arXiv preprint arXiv:2211.12588 , 2022

    Pith/arXiv arXiv 2022

  34. [34]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. arXiv preprint arXiv:2305.20050 , 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Cov: A new framework for evaluating the factuality of cot in llms

    Kanyuan Zhou, Zhipeng Li, Xiaoya Li, Yiquan Wang, and Jing Wang. Cov: A new framework for evaluating the factuality of cot in llms. arXiv preprint arXiv:2405.08745, 2024

    work page arXiv 2024

  36. [36]

    Towards anti-interference human activity recog- nition based on wifi subcarrier correlation selection

    Jinyang Huang, Bin Liu, Chao Chen, Hongxin Jin, Zhiqiang Liu, Chi Zhang, and Nenghai Yu. Towards anti-interference human activity recog- nition based on wifi subcarrier correlation selection. IEEE Transactions on Vehicular Technology, 69(6):6739–6754, 2020

    work page 2020

  37. [37]

    Benchmarking micro-action recognition: Dataset, methods, and applications

    Dan Guo, Kun Li, Bin Hu, Yan Zhang, and Meng Wang. Benchmarking micro-action recognition: Dataset, methods, and applications. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6238– 6252, 2024

    work page 2024

  38. [38]

    Multistability of Bi-Reaction Networks

    Lifan Zheng, Yixin Zhang, Hong-Yu Qu, Zhen-Dong Lin, Wayne Xin Zhang, Huachun Chen, Xi Chen, Shijin Zhao, Jing Liu, Enhong Chen, et al. A survey on chain of thought reasoning: Advances, frontiers and future. arXiv preprint arXiv:2405.05103 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    Unipotent quantum coordinate ring and cominuscule prefundamental representations

    Jie Huang, Siliang Gu, Linyi Gong, Yuan Zhao, Yilun Liu, Yuxi Jiang, Shuyuan Lu, Yudi Li, Yase Zhang, Zike Liu, et al. Thinking about thinking: A survey on large language model reasoning. arXiv preprint arXiv:2406.02870, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  40. [40]

    HOIN: High-Order Implicit Neural Representations

    Yuxuan Xie, Zhaowei Zhang, Yijia Wang, Yang Liu, and Ming Li. Characterizing and mitigating the coherence gap in chain-of-thought reasoning. arXiv preprint arXiv:2404.14674 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  41. [41]

    Ising model on a Galton-Watson tree with a sparse random external field

    Zhaofeng Feng, He Zhang, Zhaoxuan Chen, Xiaoya Li, Kanyuan Zhou, Zhipeng Li, Jing Wang, and Yiquan Wang. A causal analysis of the role of chain-of-thought in llms. arXiv preprint arXiv:2310.09169 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    Attention-based gesture recognition using commodity wifi devices

    Yu Gu, Huan Yan, Xiang Zhang, Yantong Wang, Jinyang Huang, Yusheng Ji, and Fuji Ren. Attention-based gesture recognition using commodity wifi devices. IEEE Sensors Journal, 23(9):9685–9696, 2023

    work page 2023

  43. [43]

    Balancing Test Accuracy and Security in Computerized Adaptive Testing

    Zhaowei He, Zhaowei Zhang, Yijia Wang, Yang Liu, and Ming Li. Llm- viz: A visual analytics system for post-hoc analysis of language models. arXiv preprint arXiv:2305.18312 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  44. [44]

    Classical Mechanics in Noncommutative Spaces: Confinement and More

    Xinyuan Chen, Zifan Zhang, Yuxin Wang, Ke Zhang, Yiran Wang, Qi Wang, and Kwan-Liu Wang. Visar: A human-ai loop for reasoning- aware visual investigation of llms. arXiv preprint arXiv:2405.09348 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  45. [45]

    Integrated Sensing and Communications for IoT: Synergies with Key 6G Technology Enablers

    Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. Chain-of-thought hub: A nexus of cot reasoning traces. arXiv preprint arXiv:2309.13542, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  46. [46]

    Transformerlens: An open-source library for mechanistic interpretability

    Neel Nanda and Joseph Bloom. Transformerlens: An open-source library for mechanistic interpretability. https://github.com/neelnanda-io/ TransformerLens, 2022

    work page 2022

  47. [47]

    Neurosurgeon: A toolkit for sub-network analysis and manipulation in neural networks

    Jonathan Fan, Ziyi Liu, Zhaozhuo Wang, Yun-Hsuan Chen, Ju-Chieh Lee, Tianshi Li, Yilun Wang, and Yida Wang. Neurosurgeon: A toolkit for sub-network analysis and manipulation in neural networks. In 2021 IEEE Visualization Conference (VIS) , pages 151–155. IEEE, 2021

    work page 2021

  48. [48]

    Human-in-the-loop for data collection: A case study on abstractive summarization

    Yang Gao, De-Chuan Dou, Junjie Li, Yujing Li, and Ru-Yuan Zhang. Human-in-the-loop for data collection: A case study on abstractive summarization. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1737–1747, 2022

    work page 2022

  49. [49]

    Improving Generalization in Language Model-Based Text-to-SQL Semantic Parsing: Two Simple Semantic Boundary-Based Techniques

    Boyuan Zhao, Ruijie Zhang, Shuo Chang, and Yang Liu. Iterative-cot: An iterative human-in-the-loop framework for reasoning. arXiv preprint arXiv:2305.17378, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  50. [50]

    Inter- active corpus exploration for generalizing from a single example

    Piyawat Lertvittayakumjorn, Hyesu Kim, and Francesca Toni. Inter- active corpus exploration for generalizing from a single example. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1041–1055, 2023

    work page 2023

  51. [51]

    Improving factuality and reasoning in language models through multi-agent debate

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multi-agent debate. arXiv preprint arXiv:2305.14325 , 2023

    Pith/arXiv arXiv 2023

  52. [52]

    Few-Shot Object Detection with Sparse Context Transformers

    Peifeng Li and Hai Zhao. Chainpoll: A language-model-based polling system for reasoning. arXiv preprint arXiv:2402.09315 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  53. [53]

    Editing large language models: A new research direction

    Yun-Zhu Zhang, Yu-An Yao, Zhi-Ting Chen, Bo Li, De-Chuan Peng, Jian-Jun Huang, Yu-Ting Zhuang, and Min-Ling Chen. Editing large language models: A new research direction. arXiv preprint arXiv:2401.01286, 2024

    Pith/arXiv arXiv 2024

  54. [54]

    Lo- cating and editing factual associations in gpt

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Lo- cating and editing factual associations in gpt. Advances in Neural Information Processing Systems , 35:17359–17372, 2022

    work page 2022

  55. [55]

    Plug-and-Play Posterior Sampling under Mismatched Measurement and Prior Models

    Yuval Pinter, Ryan Ott, and Yonatan Belinkov. Editing models with task-specific prompts. arXiv preprint arXiv:2310.03546 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  56. [56]

    Doshi, D

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Philliber, Richard Chen, Nathaniel Foote, Esin R-M, A. Doshi, D. Drain, et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405 , 2023

    Pith/arXiv arXiv 2023

  57. [57]

    G-NeRF: Geometry-enhanced Novel View Synthesis from Single-View Images

    Alex Turner, Ameya Variengien, Monte Conerly, Leo Smith, and Jacob Steinhardt. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2404.07474, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  58. [58]

    Critic: Large language models can self-correct with tool-interactive critiquing

    Zhibin Gou, Zhihong Qin, Linyang Ge, Runxin Zhang, Yapei Chen, Simin Shi, Weize Wang, Yong Cheng, Jian-Sun Zhang, and Ming Chen. Critic: Large language models can self-correct with tool-interactive critiquing. Transactions on Machine Learning Research , 2023

    work page 2023

  59. [59]

    Statistical curriculum learning: An elimination algorithm achieving an oracle risk

    Ruochen Zhao, Yutao Zhou, Xinyi Zhang, and Yue Zhang. Verify-and- edit: A knowledge-enhanced chain-of-thought framework. arXiv preprint arXiv:2402.13366, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  60. [60]

    Un- veiling the ”dark matter” in llms’ reasoning: A survey on interpretability of chain-of-thought

    Yijia Wang, Zhaowei Zhang, Zhaowei He, Yang Liu, and Ming Li. Un- veiling the ”dark matter” in llms’ reasoning: A survey on interpretability of chain-of-thought. arXiv preprint arXiv:2310.01174 , 2023

    Pith/arXiv arXiv 2023

  61. [61]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Kevin Hall, Luyu Gao, Sreyan Majumder, Julian McAuley, Peter Clark, and Sean Welleck. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems , 36, 2023

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.