arxiv: 2605.07315 · v1 · submitted 2026-05-08 · 💻 cs.CL

Recognition: no theorem link

LaTER: Efficient Test-Time Reasoning via Latent Exploration and Explicit Verification

Xuan Li , Yining Wang , Yuchen Liu , Guanjun Liu , Delai Qiu , Shengping Liu , Jiaen Liang , Wei Huang

show 2 more authors

Jun Yu Junnan Zhu

Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:59 UTC · model grok-4.3

classification 💻 cs.CL

keywords latent reasoningchain-of-thoughttest-time efficiencyLLM inference optimizationhidden state projectionentropy-based switchingAIME benchmarkfine-tuned reasoning

0 comments

The pith

LaTER lets models explore reasoning in continuous latent space then switch to explicit verification, cutting tokens 16-32% while matching or raising accuracy.

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

Chain-of-thought forces every intermediate step into discrete tokens, raising inference cost. Pure latent reasoning can bypass the verification steps that symbolic tasks require. LaTER splits the work into a bounded latent exploration phase followed by an explicit chain-of-thought verification phase. The switch is triggered by projecting final-layer hidden states back into the input embedding space and checking entropy together with the model's own stop-token signals. If this interface works, models can finish hard problems with substantially fewer generated tokens and without loss of correctness.

Core claim

Strong reasoning models already exhibit structured trajectories in latent space. Projecting the final-layer hidden states into the input embedding space while preserving the KV cache allows the model to continue from that trajectory into explicit tokens. Entropy measurements and native stop-token predictions identify when the latent phase has produced a usable intuition, triggering a safe handoff to verification. On Qwen3-14B the training-free version reduces tokens 16-32% across benchmarks while matching or improving accuracy; fine-tuning on Latent-Switch-69K further raises AIME 2025 accuracy to 80% with 33% fewer tokens than standard chain-of-thought.

What carries the argument

The LaTER two-stage switch: bounded latent exploration via hidden-state projection and KV-cache reuse, followed by explicit CoT triggered by entropy and stop-token probes.

If this is right

Training-free LaTER reduces total token usage by 16%-32% on several benchmarks while matching or improving accuracy.
On Qwen3-14B it improves AIME 2025 from 70.0% to 73.3% accuracy while lowering tokens from 15,730 to 10,661.
Fine-tuned LaTER reaches 80.0% accuracy on AIME 2025 using 33% fewer tokens than standard CoT.
Strong reasoning models produce structured latent trajectories that the projection interface can access.

Where Pith is reading between the lines

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

Latent space appears to hold condensed solution intuitions that explicit verification can confirm or correct.
The switch detection could be made adaptive to problem difficulty, allocating more latent steps to harder instances.
The same latent-to-explicit handoff might combine with other inference techniques such as speculative decoding for additional savings.

Load-bearing premise

Projecting final-layer hidden states back to the input embedding space and using entropy plus model-native stop-token probes can reliably detect structured latent trajectories and trigger a safe switch to explicit verification without losing essential reasoning content.

What would settle it

Applying the training-free LaTER procedure to Qwen3-14B on AIME 2025 and finding that token counts remain at or above the 15,730 baseline while accuracy stays at or below 70% would show that the claimed efficiency gain does not occur.

Figures

Figures reproduced from arXiv: 2605.07315 by Delai Qiu, Guanjun Liu, Jiaen Liang, Junnan Zhu, Jun Yu, Shengping Liu, Wei Huang, Xuan Li, Yining Wang, Yuchen Liu.

**Figure 1.** Figure 1: Overview of training-free LaTER. Given a user prompt, the model first enters a latent reasoning phase, where the final-layer hidden state is projected back into the input embedding space and reused as the next-step input, without committing to visible tokens. The model then switches to explicit CoT decoding, reusing the latent KV cache to generate reasoning steps and the final answer. The probe token yˆs i… view at source ↗

**Figure 2.** Figure 2: Entropy over normalized reasoning progress on AIME 2025 for Qwen3-14B. Blue: mean latent-reasoning entropy after aligning each example from latent start to end. Red: mean CoT entropy after normalizing each sentence by within-sentence progress. Phenomenon 1: probe tokens reveal autoregressive stopping structure. Early latent states often decode to low-content probes, such as empty strings or repeated new… view at source ↗

**Figure 3.** Figure 3: Accuracy and token usage on AIME 2025 as the fixed latent-step budget varies. The fixed-step curve is non-monotonic: performance first improves as the latent budget increases, then degrades when the return to explicit reasoning is delayed too long. This pattern supports the role separation behind LaTER. Latent exploration is useful up to a point, but difficult problems still benefit from an explicit symb… view at source ↗

**Figure 5.** Figure 5: PCA trajectories of Qwen3-14B on an AIME 2025 example under adaptive switching. The [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 4.** Figure 4: Effect of the entropy threshold τH on training-free adaptive LaTER for Qwen3- 14B on AIME 2025 Why does adaptive switching help? A fixed horizon gives every instance the same latent budget, regardless of difficulty or internal confidence. Figures 3 and Figures 4 together show that neither too few nor too many latent steps are desirable. If the latent phase is too short, the model leaves latent exploration… view at source ↗

**Figure 6.** Figure 6: Token statistics of the distilled corpus. The figure compares original and distilled reasoning [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

**Figure 7.** Figure 7: Overview of training pipeline. We construct latent-reasoning training sequences with latent and explicit reasoning segments, train the model with supervised and teacher-matching objectives, and obtain a model that first performs latent reasoning and then switches to explicit CoT generation [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

**Figure 8.** Figure 8: Per-step entropy distributions during the training-free latent phase on AIME 2025. For [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

**Figure 9.** Figure 9: Composition of the Latent-Switch-69K training corpus after filtering and distillation. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗

**Figure 10.** Figure 10: Prompt for math questions. Prompt Target Question: {question} You are a helpful assistant. You must reason step-by-step to solve the provided **Target Question** without outputting other irrelevant information. Your final answer must be selected from A, B, C, D, ... For example \boxed{A}. Do not add any other contents inside the box. Now, reason step by step and output the final answer inside \boxed{YOUR_… view at source ↗

**Figure 11.** Figure 11: Prompt for multi-choice questions. Prompt Target Question: {question} You must put all python code as self-contained Python function(s) in markdown code blocks. For example: ```python import math def add(a, b): return a + b ``` Do not add any other contents inside the markdown code block. Now, reason step by step and output the final answer [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗

**Figure 12.** Figure 12: Prompt for code problems. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗

**Figure 13.** Figure 13: System prompt to generate solution intuition. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗

**Figure 14.** Figure 14: User prompt to generate solution intuition. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗

**Figure 15.** Figure 15: System prompt to generate shorter CoT. Prompt Solve the following problem by continuing from your intuition. Your intuition may be correct or incorrect. Do not ignore it. Continue from it and finish the solution. REMEMBER: The intuition comes from your previous conversation with yourself. It is not the user’s intuition. Problem prompt: <<<PROMPT>>> {prompt} <<<END_PROMPT>>> Your Intuition: <<<INTUITION>>>… view at source ↗

**Figure 16.** Figure 16: User prompt to generate shorter CoT. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗

**Figure 17.** Figure 17: Low-entropy latent reasoning can still fail. On AIME 2025 Problem 10, the baseline [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗

read the original abstract

Chain-of-thought (CoT) reasoning improves large language models (LLMs) on difficult tasks, but it also makes inference expensive because every intermediate step must be generated as a discrete token. Latent reasoning reduces visible token generation by propagating continuous states, yet replacing explicit derivations with latent computation can hurt tasks that require symbolic checking. We propose Latent-Then-Explicit Reasoning (LaTER), a two-stage paradigm that first performs bounded exploration in a continuous latent space and then switches to explicit CoT for verification and answer generation. In a training-free instantiation, LaTER projects final-layer hidden states back to the input embedding space, preserves the latent KV cache, and uses entropy and model-native stop-token probes to decide when to switch. We find that strong reasoning models already exhibit structured latent trajectories under this interface. On Qwen3-14B, training-free LaTER reduces total token usage by 16%-32% on several benchmarks while matching or improving accuracy on most of them; for example, it improves AIME 2025 from 70.0% to 73.3% while reducing tokens from 15,730 to 10,661. We further construct Latent-Switch-69K, a supervised corpus that pairs condensed solution intuitions with shortened explicit derivations. Fine-tuning with latent rollout and halting supervision yields additional gains: trained LaTER reaches 80.0% accuracy on AIME 2025, 10.0 points above the standard CoT baseline, while using 33% fewer tokens. Our code, data, and model are available at https://github.com/TioeAre/LaTER.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LaTER gives a practical training-free way to mix latent exploration with explicit CoT and cut tokens on reasoning benchmarks, but the entropy-based switch from projected hidden states looks under-tested.

read the letter

The core contribution is a two-stage setup: bounded latent work in hidden states, then a switch to token generation for verification. In the training-free version they project final-layer states back to the embedding space, keep the KV cache intact, and trigger the switch with entropy plus the model's own stop-token signals. On Qwen3-14B this trims 16-32% tokens across benchmarks while matching or beating standard CoT accuracy, with a fine-tuned version hitting 80% on AIME 2025 at 33% fewer tokens. They also ship Latent-Switch-69K and the code, which is straightforward to check.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes LaTER, a two-stage reasoning paradigm for LLMs: bounded exploration in continuous latent space (via projection of final-layer hidden states back to input embeddings while preserving KV cache) followed by explicit CoT verification triggered by entropy and model-native stop-token probes. Training-free LaTER on Qwen3-14B reduces token usage by 16-32% on benchmarks while matching or improving accuracy (e.g., AIME 2025: 70.0% to 73.3%, tokens 15,730 to 10,661). Fine-tuning on the new Latent-Switch-69K corpus yields further gains (80.0% on AIME 2025 with 33% fewer tokens than standard CoT). Code, data, and models are released.

Significance. If the latent-to-explicit switch reliably preserves reasoning content, LaTER could meaningfully advance efficient test-time reasoning by exploiting structured latent trajectories in strong models without full token generation. The public release of code, Latent-Switch-69K data, and models is a clear strength supporting reproducibility. The reported efficiency-accuracy trade-offs are practically relevant for long-horizon tasks, but hinge on unverified assumptions about probe robustness.

major comments (2)

[§3 (Methods)] §3 (Methods): The training-free LaTER claims rest on the entropy and stop-token probes (applied to projected hidden states) safely detecting usable latent structure before switching. No ablation on probe thresholds, no qualitative inspection of projected tokens or trajectories, and no failure-case analysis on long-horizon tasks (e.g., AIME) are provided. This leaves open whether the 16-32% token reductions and accuracy lifts (e.g., +3.3 on AIME) arise from genuine latent reasoning or simply shorter sampling.
[§4 (Experiments)] §4 (Experiments): Accuracy improvements such as the 3.3-point AIME 2025 gain and the fine-tuned 10-point lift are reported without error bars, multiple random seeds, statistical significance tests, or details on run variance. This undermines assessment of whether gains are robust or affected by benchmark selection/post-hoc choices, especially given the abstract's limited methodological visibility.

minor comments (2)

[Abstract and §4] Abstract and §4: Explicitly state the number of evaluation runs or variance for all reported metrics to aid interpretation.
[Throughout] Throughout: Ensure any figures depicting latent trajectories or probe decisions include detailed captions and axis labels for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below and describe the revisions we will make to strengthen the paper.

read point-by-point responses

Referee: [§3 (Methods)] §3 (Methods): The training-free LaTER claims rest on the entropy and stop-token probes (applied to projected hidden states) safely detecting usable latent structure before switching. No ablation on probe thresholds, no qualitative inspection of projected tokens or trajectories, and no failure-case analysis on long-horizon tasks (e.g., AIME) are provided. This leaves open whether the 16-32% token reductions and accuracy lifts (e.g., +3.3 on AIME) arise from genuine latent reasoning or simply shorter sampling.

Authors: We acknowledge that the current version lacks ablations on probe thresholds, qualitative trajectory inspections, and explicit failure-case analysis. In the revision we will add a dedicated subsection with threshold ablations (varying entropy and stop-token cutoffs) and report resulting accuracy/token trade-offs on AIME and other benchmarks. We will also include qualitative examples of projected tokens and latent trajectories for selected AIME problems, plus a short failure-case study showing when the switch occurs too early or too late. These additions will directly test whether gains derive from structured latent reasoning rather than reduced sampling length. revision: yes
Referee: [§4 (Experiments)] §4 (Experiments): Accuracy improvements such as the 3.3-point AIME 2025 gain and the fine-tuned 10-point lift are reported without error bars, multiple random seeds, statistical significance tests, or details on run variance. This undermines assessment of whether gains are robust or affected by benchmark selection/post-hoc choices, especially given the abstract's limited methodological visibility.

Authors: We agree that the absence of error bars, multi-seed statistics, and significance tests limits evaluation of robustness. In the revised manuscript we will rerun the main experiments (training-free and fine-tuned) with five random seeds, report mean and standard deviation for accuracy and token counts, and include paired statistical tests (e.g., Wilcoxon signed-rank) for the reported gains. We will also expand the abstract and experimental setup description to improve methodological visibility. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical head-to-head evaluation on public benchmarks with no self-referential equations or fitted predictions

full rationale

The paper introduces LaTER as a two-stage procedure (latent exploration then explicit verification) whose training-free version projects final-layer states, preserves KV cache, and switches via entropy plus native stop-token probes. All headline results—16-32% token reduction and accuracy gains such as AIME 2025 from 70.0% to 73.3%—are obtained by direct comparison against standard CoT on fixed public benchmarks. No equation or derivation reduces the reported token counts or accuracies to parameters whose values are defined by those same metrics. The observation that strong models exhibit structured latent trajectories is presented as an empirical finding, not as an input that is re-derived by construction. No self-citation chain, ansatz smuggling, or renaming of known results is used to justify the central claims. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities beyond standard LLM components; the method implicitly assumes that LLM hidden states contain explorable structured trajectories for reasoning tasks.

axioms (1)

domain assumption Strong reasoning models exhibit structured latent trajectories in final-layer hidden states that can be projected and probed for switching decisions.
This premise underpins the bounded latent exploration and the entropy/stop-token switch mechanism described in the abstract.

pith-pipeline@v0.9.0 · 5623 in / 1336 out tokens · 40118 ms · 2026-05-11T01:59:35.244402+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 8 internal anchors

[1]

Chain-of-thought prompting elicits reason- ing in large language models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reason- ing in large language models. InAdvances in Neural Information Processing Sys- tems, 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/ file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf

work page 2022
[2]

Available: http://dx.doi.org/10.1038/s41586-025-09422-z

Daya Guo, Dejian Yang, Haowei Zhang, et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.Nature, September 2025. doi: 10.1038/s41586-025-09422-z. URLhttp://dx.doi.org/10.1038/s41586-025-09422-z

work page doi:10.1038/s41586-025-09422-z 2025
[3]

Attention is all you need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Informa- tion Processing Systems, 2017. URL https://proceedings.neurips.cc/paper_files/ paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf

work page 2017
[4]

Training large language models to reason in a continuous latent space

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason E Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. InSecond Conference on Language Modeling, 2025. URL https://openreview.net/forum?id= Itxz7S4Ip3

work page 2025
[6]

arXiv preprint arXiv:2511.20639 , year =

Jiaru Zou, Xiyuan Yang, Ruizhong Qiu, Gaotang Li, Katherine Tieu, Pan Lu, Ke Shen, Hang- hang Tong, Yejin Choi, Jingrui He, James Zou, Mengdi Wang, and Ling Yang. Latent collabora- tion in multi-agent systems, 2025. URLhttps://arxiv.org/abs/2511.20639

work page arXiv 2025
[7]

SoftCoT: Soft chain-of-thought for efficient reasoning with LLMs

Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. SoftCoT: Soft chain-of-thought for efficient reasoning with LLMs. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vienna, Austria, July 2025. doi: 10.18653/ v1/2025.acl-long.1137. URLhttps://aclanthology.org/2025.acl-long.1137/

work page 2025
[8]

The latent space: Foundation, evolution, mechanism, ability, and outlook,

Xinlei Yu, Zhangquan Chen, Yongbo He, Tianyu Fu, Cheng Yang, Chengming Xu, Yue Ma, Xiaobin Hu, Zhe Cao, Jie Xu, Guibin Zhang, Jiale Tao, Jiayi Zhang, Siyuan Ma, Kaituo Feng, Haojie Huang, Youxing Li, Ronghao Chen, Huacan Wang, Chenglin Wu, Zikun Su, Xiaogang Xu, Kelu Yao, Kun Wang, Chen Gao, Yue Liao, Ruqi Huang, Tao Jin, Cheng Tan, Jiangning Zhang, Wenqi...

work page
[9]

URLhttps://arxiv.org/abs/2604.02029

work page arXiv
[10]

A survey on latent reasoning.arXiv preprint arXiv:2507.06203, 2025

Rui-Jie Zhu, Tianhao Peng, Tianhao Cheng, Xingwei Qu, Jinfa Huang, Dawei Zhu, Hao Wang, Kaiwen Xue, Xuanliang Zhang, Yong Shan, Tianle Cai, Taylor Kergan, Assel Kembay, Andrew Smith, Chenghua Lin, Binh Nguyen, Yuqi Pan, Yuhong Chou, Zefan Cai, Zhenhe Wu, Yongchi Zhao, Tianyu Liu, Jian Yang, Wangchunshu Zhou, Chujie Zheng, Chongxuan Li, Yuyin Zhou, Zhoujun...

work page arXiv 2025
[11]

Latent reasoning in llms as a vocabulary-space superposition,

Jingcheng Deng, Liang Pang, Zihao Wei, Shicheng Xu, Zenghao Duan, Kun Xu, Yang Song, Huawei Shen, and Xueqi Cheng. Latent reasoning in llms as a vocabulary-space superposition,

work page
[12]

URLhttps://arxiv.org/abs/2510.15522v1. 10

work page arXiv
[13]

Think silently, think fast: Dynamic latent compression of LLM reasoning chains

Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Ruihua Song, and Jian Luan. Think silently, think fast: Dynamic latent compression of LLM reasoning chains. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URL https://openreview. net/forum?id=AQsko3PPUe

work page 2026
[14]

CODI: Com- pressing chain-of-thought into continuous space via self-distillation

Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. CODI: Com- pressing chain-of-thought into continuous space via self-distillation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Suzhou, China, November

work page 2025
[15]

CODI: compress- ing chain-of-thought into continuous space via self-distillation

doi: 10.18653/v1/2025.emnlp-main.36. URL https://aclanthology.org/2025. emnlp-main.36/

work page doi:10.18653/v1/2025.emnlp-main.36 2025
[16]

Qwen3 Technical Report

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, et al. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388

work page internal anchor Pith review Pith/arXiv arXiv 2025
[17]

Olmo 3

Team Olmo, Allyson Ettinger, Amanda Bertsch, et al. Olmo 3, 2026. URL https://arxiv. org/abs/2512.13961

work page internal anchor Pith review Pith/arXiv arXiv 2026
[18]

Matharena: Evaluating llms on uncontaminated math competitions, February 2025

Mislav Balunovi ´c, Jasper Dekoninck, Ivo Petrov, Nikola Jovanovi ´c, and Martin Vechev. Matharena: Evaluating llms on uncontaminated math competitions, February 2025. URL https://matharena.ai/

work page 2025
[19]

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

work page internal anchor Pith review Pith/arXiv arXiv 2023
[20]

Training Verifiers to Solve Math Word Problems

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. URL https://arxiv.org/ abs/2110.14168

work page internal anchor Pith review Pith/arXiv arXiv 2021
[21]

David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level google- proof q&a benchmark. InFirst Conference on Language Modeling, 2024. URL https: //openreview.net/forum?id=Ti67584b98

work page 2024
[22]

Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge.arXiv:1803.05457v1, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018
[23]

Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation

Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https: //openreview.net/forum?id=1qvx610Cu7

work page 2023
[24]

Evaluating language models for efficient code generation

Jiawei Liu, Songrun Xie, Junhao Wang, Yuxiang Wei, Yifeng Ding, and Lingming Zhang. Evaluating language models for efficient code generation. InFirst Conference on Language Modeling, 2024. URLhttps://openreview.net/forum?id=IBCBMeAhmC

work page 2024
[25]

Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models, 2026. URLhttps://arxiv.org/abs/2601.18734

work page internal anchor Pith review arXiv 2026
[26]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019. URL https://openreview.net/forum? id=Bkg6RiCqY7

work page 2019
[27]

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. URLhttps://arxiv.org/abs/2307.08691

work page internal anchor Pith review arXiv 2023
[28]

Zero: memory optimiza- tions toward training trillion parameter models

Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: memory optimiza- tions toward training trillion parameter models. InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’20. IEEE Press, 2020. 11

work page 2020
[29]

Soft thinking: Unlocking the reasoning potential of LLMs in continuous concept space

Zhen Zhang, Xuehai He, Weixiang Yan, Ao Shen, Chenyang Zhao, and Xin Eric Wang. Soft thinking: Unlocking the reasoning potential of LLMs in continuous concept space. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URL https://openreview.net/forum?id=ByQdHPGKgU

work page 2026
[30]

Swireasoning: Switch-thinking in latent and explicit for pareto-superior reasoning llms,

Dachuan Shi, Abedelkadir Asi, Keying Li, Xiangchi Yuan, Leyan Pan, Wenke Lee, and Wen Xiao. Swireasoning: Switch-thinking in latent and explicit for pareto-superior reasoning llms,

work page
[31]

URLhttps://arxiv.org/abs/2510.05069

work page arXiv
[32]

SeLaR: Selective Latent Reasoning in Large Language Models

Renyu Fu and Guibo Luo. Selar: Selective latent reasoning in large language models, 2026. URLhttps://arxiv.org/abs/2604.08299. 12 A Per-step entropy heterogeneity in training-free latent reasoning We provide a finer-grained view of the stopping signals used by the training-free version of LaTER. For each AIME 2025 problem, we begin with latent reasoning a...

work page internal anchor Pith review Pith/arXiv arXiv 2026