Writing study: ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Akshat Ramachandran, Marina Neseem, Charbel Sakr, Rangharajan Venkatesan, Brucek Khailany, Tushar Krishna · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Inference, algorithms & systems efficiency (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

35 physical pages

Before references / appendix

≈ 10.2 page-equivalents

Main figures / tables

11 / 5

Attached appendix figures / tables

7 / 10

Attached appendix span

≈ 20.31 page-equivalents

Before references / appendix: ≈ 29.14%References / intervening material: ≈ 12.83%Attached appendix: ≈ 58.03%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: ThinKV: restore Figure 2 on physical page 3, Table 2 on page 8, and Figure 11 on page 10. All three are real side-wrapped captions whose labels and bodies were split or mixed with body text by Poppler. Each location was visually checked.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 1.27 pp.	3.63%
2 WHY QUANTIZATION+EVICTION ?	3	≈ 0.3 pp.	0.86%
3 MOTIVATING ANALYSES	3	≈ 1.31 pp.	3.74%
4 THINKV METHODOLOGY	4	≈ 1.58 pp.	4.51%
5 THINKV SYSTEM IMPLEMENTATION	6	≈ 1.26 pp.	3.6%
6 EVALUATION	7	≈ 3.23 pp.	9.23%
7 CONCLUSION	10	≈ 0.31 pp.	0.89%
ACKNOWLEDGEMENTS	11	≈ 0.1 pp.	0.29%
REFERENCES	11	≈ 4.49 pp.	12.83%
A PPENDIX	15	≈ 1.79 pp.	5.11%
B EXTENDED RELATED WORKS	17	≈ 1.16 pp.	3.31%
C SUPPLEMENTARY BACKGROUND	18	≈ 3.09 pp.	8.83%
D SUPPLEMENTARY DETAILS ON THINKV	21	≈ 3 pp.	8.57%
E EXTENDED EVALUATIONS	24	≈ 9.9 pp.	28.29%
F LIMITATIONS	34	≈ 0.12 pp.	0.34%
G IMPACT STATEMENT	34	≈ 0.34 pp.	0.97%
H LLM USAGE STATEMENT	35	≈ 0.89 pp.	2.54%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 3
• Figure 3 · p. 3
• Figure 4 · p. 4
• Figure 5 · p. 4
• Figure 6 · p. 5
• Figure 7 · p. 6
• Figure 8 · p. 7
• Table 1 · p. 8
• Table 2 · p. 8
• Table 3 · p. 8
• Figure 10 · p. 9
• Figure 9 · p. 9
• Table 4 · p. 9
• Table 5 · p. 10
• Figure 11 · p. 10

Attached appendix

• Table 6 · p. 17
• Figure 12 · p. 21
• Table 7 · p. 21
• Figure 13 · p. 24
• Figure 14 · p. 25
• Figure 15 · p. 25
• Figure 16 · p. 26
• Table 8 · p. 26
• Table 9 · p. 26
• Table 10 · p. 26
• Table 11 · p. 27
• Figure 17 · p. 27
• Table 12 · p. 27
• Table 13 · p. 28
• Figure 18 · p. 28
• Table 14 · p. 28
• Table 15 · p. 28

Argument and organization

1. Motivating analyses

   Establish attention-sparsity regimes, counterfactual importance and inter-segment dependence before introducing compression.

2. ThinKV and Continuous Thinking

   Connect thought-adaptive precision/eviction to memory-slot reuse and runtime kernels.

3. Evaluation and supplementary details

   Separate accuracy, batch-capacity throughput, iso-batch controls, implementation and limitations.

Problem, claim and contribution

The introduction narrows the problem to long generated reasoning, distinguishing it from prompt-cache compression. Four contributions map directly to observation, algorithm and systems sections.

Sections 1–3. · Main paper

Argument and section order

Calibration, representative keyword labels and runtime classification are separated. Keywords aid interpretation only; actual classification uses attention sparsity, so the semantic names should not be mistaken for supervised thought labels.

Sections 3.1 and 4.1; Appendices D.1–D.2. · Main paper

Figures, captions and visual explanation

Figure 6 makes memory management tangible with token letters, logical/physical blocks and evolving mask rows. Its labeled walkthrough is valuable, although narrow table labels require enlargement.

Figure 6, page 5; Appendix D.6. · Main paper

Tables, comparisons and evidence

Maximum-batch throughput and iso-batch/iso-compression controls appear in separate blocks. This prevents attributing all of the headline speedup to faster single-request decoding.

Table 2 and Section 6.2, page 8. · Main paper

Appendix and reproducibility

The appendix documents proof-of-concept versus vLLM integration, an AIME mixture of 15 questions each from 2024/2025, and a one-category fallback for non-reasoning models.

Appendices E.1–E.2, E.10 and E.12. · Main paper

Clarity and writing conventions

Near-lossless must retain its configuration: the MobileLLM example drops from 67.5 to 60.1, while GPT-OSS-120B loses 1.9–2.5 points. Both are valuable boundary cases against an unconditional headline.

Appendix E.6–E.7, Tables 8–9. · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair each algorithmic compression gain with an implementation mechanism and an iso-resource control, then distinguish memory capacity from user latency.

Boundary: Attention-sparsity clusters are heuristic proxies for semantic thoughts. Compression changes the generated trajectory; KV percentage is not total-model memory, and distilled backbones do not make ThinKV a distillation algorithm.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–10: Sections 1–7; references 11–15.

Appendix / supplement

Pages 15–35: contents, Appendices A–H including full calibration/generation algorithms, attention proof, all extended evaluations and complete reasoning traces.

Visually inspected PDF pages

5, 8

• ICLR 2026 final proceedings · 35 pages