arxiv: 2605.06165 · v1 · submitted 2026-05-07 · 💻 cs.AI

Recognition: unknown

Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost

Richmond Sin Jing Xuan, Rishabh Bhardwaj, Soujanya Poria

Pith reviewed 2026-05-08 10:15 UTC · model grok-4.3

classification 💻 cs.AI

keywords post-reasoninglarge language modelsinstruction tuningreasoning efficiencydirect answerinference optimization

0 comments

The pith

Post-Reasoning improves non-thinking large language models by having them justify answers after the final response, at no extra cost.

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

Large language models often generate long reasoning traces before answering, which adds latency and cost even when not needed. This paper proposes Post-Reasoning, where the model first gives the answer and then explains how it arrived at it. The method boosts accuracy on many reasoning and knowledge benchmarks in most tested cases, with an average relative gain of over 17 percent. A training version that internalizes this behavior pushes performance higher still. The result is better direct-answer performance without paying for extra tokens at inference time.

Core claim

Post-Reasoning conditions models to generate justifications after the final answer rather than before. This simple change improves performance in over 88 percent of 117 model-benchmark combinations, delivering a 17.37 percent mean relative improvement. Supervised post-reason tuning internalizes the behavior and adds another 8 percent on average. The approach sets a new upper limit on what direct-answer models can achieve.

What carries the argument

The post-reasoning prompt structure that requires the model to output the answer first followed by its justification.

If this is right

Performance gains appear across 13 models and 9 benchmarks including math problems and general knowledge tests.
Supervised tuning on post-reasoning examples further raises accuracy beyond the prompting approach.
Final answers can be returned immediately while still benefiting from the post-generation step during training or evaluation.
Direct-answer mode becomes competitive with explicit reasoning methods at lower operational cost.

Where Pith is reading between the lines

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

The separation of answer and justification may allow models to commit to an answer before overthinking it, potentially reducing certain error patterns.
This technique could be combined with other efficiency methods to further cut inference costs.
Testing on models fine-tuned without any reasoning data might reveal whether post-reasoning can bootstrap reasoning ability from scratch.

Load-bearing premise

The observed improvements stem from the post-answer reasoning instruction rather than from changes in prompt length, specific wording, or differences in how answers are extracted.

What would settle it

Evaluating Post-Reasoning on a held-out model family and benchmark suite where accuracy stays the same or drops compared to standard direct answering.

Figures

Figures reproduced from arXiv: 2605.06165 by Richmond Sin Jing Xuan, Rishabh Bhardwaj, Soujanya Poria.

**Figure 1.** Figure 1: Meta-analysis of Post-Reasoning improvements across benchmarks. Each point denotes the relative gain over direct answering for a (model, task) pair. Gains are predominantly positive, with larger improvements on multi-step reasoning tasks (AMC, HMMT, MATH) and smaller or mixed gains on arithmetic and knowledgeintensive benchmarks. Experiments were conducted across 13 opensource and proprietary models … view at source ↗

**Figure 2.** Figure 2: Training loss curve for the Post-Reason SFT framework, demonstrating stable gradient view at source ↗

**Figure 3.** Figure 3: Extended loss convergence comparisons for all 10 Phase II models. Solid blue lines view at source ↗

**Figure 4.** Figure 4: Aggregate training loss convergence across all models on the MMLU dataset. When view at source ↗

read the original abstract

As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to inference latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose \textbf{Post-Reasoning}, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across \(117\) model--benchmark settings spanning \(13\) open and proprietary models, \(4\) model families, and \(9\) diverse reasoning and knowledge-intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over \(88.19\%\) of evaluated settings, achieving a mean relative improvements of \(17.37\%\). Furthermore, we propose supervised post-reason tuning, which further improves performance in over \(91.11\%\) of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of \(8.01\%\), demonstrating that post-reasoning can be effectively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.

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

Post-reasoning shows consistent lifts on direct-answer tasks across many models but the gains may trace to prompt elaboration rather than the answer-first order specifically.

read the letter

The main point is that telling an instruction-tuned model to output its final answer first and then justify it produces measurable gains on standard benchmarks while allowing early stopping to avoid extra tokens. The paper reports this works in 88% of 117 settings with a 17% average relative improvement, and a supervised tuning version that internalizes the behavior adds another 8% on top. That scale of testing across open and closed models plus knowledge and reasoning benchmarks is the strongest part of the work. It gives practitioners a simple augmentation that appears to raise the direct-answer ceiling without changing inference cost structure. The supervised post-reason tuning result is also useful because it shows the pattern can be learned rather than prompted every time. The soft spots are exactly where the stress-test flags them. The baseline is a plain direct prompt, so the lift could come from any added structure or length rather than the temporal reversal of answer and reasoning. No ablations are described that hold total prompt size and lexical pressure constant while only moving the reasoning block. Without those, the causal claim stays provisional. Extraction details and whether any post-processing influenced the percentages also need checking before the numbers can be taken at face value. This is for people who run LLMs in production or tune prompting pipelines and want cheap accuracy wins. A reader who cares about efficient inference will find the empirical pattern worth testing locally. It deserves peer review because the breadth of results makes the idea worth verifying, even if the mechanism needs tighter isolation. I would send it out and request the missing prompt-variant controls and extraction protocol.

Referee Report

2 major / 2 minor

Summary. The paper proposes Post-Reasoning, a prompting method that instructs LLMs to output the final answer before providing any justification. It reports empirical results across 117 model-benchmark settings (13 models, 9 benchmarks including GSM8K, GPQA, MMLU-Pro, BIG-Bench Hard) showing improvement in 88.19% of cases with mean relative gain of 17.37%, at no added inference cost. It further introduces supervised post-reason tuning that improves over the prompt baseline in 91.11% of settings by an additional 8.01% on average.

Significance. If the gains are shown to stem specifically from the answer-first ordering rather than prompt length or format artifacts, the result would be significant for inference efficiency: it suggests a way to raise the performance ceiling of direct-answer (non-CoT) inference without incurring reasoning-trace token costs. The scale of the evaluation (117 settings spanning open and proprietary models) is a clear strength and provides a broad empirical foundation.

major comments (2)

[§4] §4 (Experiments) and §4.1 (Baselines): The direct-answer baseline is compared only to the post-reasoning prompt; no ablation holds total prompt length, lexical complexity, and output-format pressure constant while varying only the temporal placement of reasoning (answer-then-justify vs. justify-then-answer vs. answer-then-any-explanation). This control is load-bearing for the central claim that the observed 17.37% mean relative improvement is caused by the post-reasoning structure itself.
[Table 1] Table 1 and §4.3 (Results): The headline 88.19% improvement rate and 17.37% mean relative gain aggregate across heterogeneous settings without reported per-benchmark or per-model-family breakdowns or statistical tests for consistency. It is therefore unclear whether the effect is broadly robust or driven by particular subsets (e.g., knowledge vs. math benchmarks).

minor comments (2)

[Abstract] The abstract states that Post-Reasoning works 'at no cost,' yet the supervised post-reason tuning variant incurs training cost; the scope of the 'no cost' claim should be clarified in the introduction and conclusion.
[§5] Figure captions and §5 (Discussion) would benefit from explicit statements of whether answer extraction and evaluation protocols were identical between baseline and post-reasoning runs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major point below and indicate where we will revise the manuscript to strengthen the evidence for our claims.

read point-by-point responses

Referee: [§4] §4 (Experiments) and §4.1 (Baselines): The direct-answer baseline is compared only to the post-reasoning prompt; no ablation holds total prompt length, lexical complexity, and output-format pressure constant while varying only the temporal placement of reasoning (answer-then-justify vs. justify-then-answer vs. answer-then-any-explanation). This control is load-bearing for the central claim that the observed 17.37% mean relative improvement is caused by the post-reasoning structure itself.

Authors: We agree that isolating the causal contribution of the answer-first ordering requires explicit controls for prompt length, lexical complexity, and output format. In the revised manuscript we will add a dedicated ablation subsection that holds these factors constant across three conditions: (i) standard direct-answer, (ii) post-reasoning (answer-then-justify), and (iii) justify-then-answer, with neutral padding text used to equalize token counts and syntactic complexity. We will also include an “answer-then-any-explanation” variant to test whether the benefit is specific to post-hoc justification. These results will be reported alongside the existing 117-setting evaluation. revision: yes
Referee: [Table 1] Table 1 and §4.3 (Results): The headline 88.19% improvement rate and 17.37% mean relative gain aggregate across heterogeneous settings without reported per-benchmark or per-model-family breakdowns or statistical tests for consistency. It is therefore unclear whether the effect is broadly robust or driven by particular subsets (e.g., knowledge vs. math benchmarks).

Authors: We acknowledge that aggregate statistics can mask heterogeneity. The revised manuscript will expand §4.3 and Table 1 with per-benchmark and per-model-family breakdowns (e.g., math vs. knowledge-intensive tasks, open vs. proprietary models). We will also add statistical tests, including the proportion of settings showing statistically significant improvement (paired Wilcoxon signed-rank test per benchmark) and consistency metrics across model families. These additions will demonstrate that the reported gains are not driven by a small subset of settings. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical prompt-augmentation study with external benchmarks

full rationale

The paper introduces Post-Reasoning as a prompt template that elicits final answers before justifications, then measures accuracy on 117 independent model-benchmark pairs (AMC, GSM8K, GPQA, MMLU-Pro, etc.). No equations, fitted parameters, uniqueness theorems, or self-citations are invoked to derive the performance numbers; the reported 88.19% improvement rate and 17.37% mean relative gain are direct empirical observations against fixed external test sets. The method is self-contained and falsifiable without reducing any claimed result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that LLM behavior can be reliably altered by prompt ordering and that the reported improvements generalize beyond the tested models and benchmarks.

axioms (1)

domain assumption LLM output order can be controlled by instruction without changing the underlying model weights or decoding parameters.
Invoked throughout the method description to justify that answer-first generation is feasible.

pith-pipeline@v0.9.0 · 5564 in / 1080 out tokens · 28496 ms · 2026-05-08T10:15:59.968409+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

299 extracted references · 147 canonical work pages · 30 internal anchors

[1]

2025 , url =

Aubakirova, Malika and Atallah, Alex and Clark, Chris and Summerville, Justin and Midha, Anjney , title =. 2025 , url =

2025
[2]

2026 , howpublished =

Deloitte's Enterprise AI Infrastructure Survey: A 2028 Outlook , author =. 2026 , howpublished =

2028
[3]

2025 , howpublished =

The State of Enterprise AI: 2025 Report , author =. 2025 , howpublished =

2025
[4]

2025 , eprint=

Qwen3 Technical Report , author=. 2025 , eprint=

2025
[5]

2025 , eprint=

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author=. 2025 , eprint=

2025
[6]

2025 , eprint=

gpt-oss-120b & gpt-oss-20b Model Card , author=. 2025 , eprint=

2025
[7]

Advances in neural information processing systems , volume=

Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=
[8]

2025 , url=

Global AI Adoption 2025 , author=. 2025 , url=

2025
[9]

Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models

Stop overthinking: A survey on efficient reasoning for large language models , author=. arXiv preprint arXiv:2503.16419 , year=

work page internal anchor Pith review arXiv
[10]

Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models

Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models , author=. arXiv preprint arXiv:2507.04023 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[11]

arXiv preprint arXiv:2604.02155 , year=

Brief Is Better: Non-Monotonic Chain-of-Thought Budget Effects in Function-Calling Language Agents , author=. arXiv preprint arXiv:2604.02155 , year=

work page arXiv
[12]

arXiv preprint arXiv:2508.13141 , year=

Optimalthinkingbench: Evaluating over and underthinking in llms , author=. arXiv preprint arXiv:2508.13141 , year=

work page arXiv
[13]

The Rise of AI in the Workplace , year =
[14]

2023 , eprint=

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. 2023 , eprint=

2023
[15]

2020 , eprint=

Language Models are Few-Shot Learners , author=. 2020 , eprint=

2020
[16]

2023 , eprint=

How Language Model Hallucinations Can Snowball , author=. 2023 , eprint=

2023
[17]

Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus

Ji, Ziwei and Lee, Nayeon and Frieske, Rita and Yu, Tiezheng and Su, Dan and Xu, Yan and Ishii, Etsuko and Bang, Ye Jin and Madotto, Andrea and Fung, Pascale , year=. Survey of Hallucination in Natural Language Generation , volume=. ACM Computing Surveys , publisher=. doi:10.1145/3571730 , number=

work page doi:10.1145/3571730
[18]

2022 , eprint=

Training language models to follow instructions with human feedback , author=. 2022 , eprint=

2022
[19]

2024 , eprint=

Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. 2024 , eprint=

2024
[20]

2022 , eprint=

Scaling Laws for Reward Model Overoptimization , author=. 2022 , eprint=

2022
[21]

2023 , eprint=

Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback , author=. 2023 , eprint=

2023
[22]

2022 , eprint=

STaR: Bootstrapping Reasoning With Reasoning , author=. 2022 , eprint=

2022
[23]

2022 , eprint=

Large Language Models Can Self-Improve , author=. 2022 , eprint=

2022
[24]

Hugging Face repository , howpublished =

Jia LI and Edward Beeching and Lewis Tunstall and Ben Lipkin and Roman Soletskyi and Shengyi Costa Huang and Kashif Rasul and Longhui Yu and Albert Jiang and Ziju Shen and Zihan Qin and Bin Dong and Li Zhou and Yann Fleureau and Guillaume Lample and Stanislas Polu , title =. Hugging Face repository , howpublished =. 2024 , publisher =

2024
[25]

2024 , eprint=

The Llama 3 Herd of Models , author=. 2024 , eprint=

2024
[26]

2025 , eprint=

Gemma 3 Technical Report , author=. 2025 , eprint=

2025
[27]

arXiv , year=

Mistral 7B , author=. arXiv , year=
[28]

2021 , eprint=

LoRA: Low-Rank Adaptation of Large Language Models , author=. 2021 , eprint=

2021
[29]

2023 , eprint=

Parameter-Efficient Fine-Tuning Methods for Pretrained Language Models: A Critical Review and Assessment , author=. 2023 , eprint=

2023
[30]

2021 , eprint=

Measuring Mathematical Problem Solving With the MATH Dataset , author=. 2021 , eprint=

2021
[31]

2021 , eprint=

Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint=

2021
[32]

2023 , eprint=

GPQA: A Graduate-Level Google-Proof Q&A Benchmark , author=. 2023 , eprint=

2023
[33]

2024 , eprint=

MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark , author=. 2024 , eprint=

2024
[34]

Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them

Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them , author=. arXiv preprint arXiv:2210.09261 , year=

work page internal anchor Pith review arXiv
[35]

arXiv preprint arXiv:2409.18433 , year=

Easy2Hard-Bench: Standardized Difficulty Labels for Profiling LLM Performance and Generalization , author=. arXiv preprint arXiv:2409.18433 , year=

work page arXiv
[36]

2026 , eprint=

Ministral 3 , author=. 2026 , eprint=

2026
[37]

2025 , eprint=

Gemini: A Family of Highly Capable Multimodal Models , author=. 2025 , eprint=

2025
[38]

2024 , eprint=

Orca-Math: Unlocking the potential of SLMs in Grade School Math , author=. 2024 , eprint=

2024
[39]

2022 , eprint=

Emergent Abilities of Large Language Models , author=. 2022 , eprint=

2022
[40]

2024 , eprint=

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters , author=. 2024 , eprint=

2024
[41]

2017 , eprint=

SGDR: Stochastic Gradient Descent with Warm Restarts , author=. 2017 , eprint=

2017
[42]

2021 , eprint=

A General Language Assistant as a Laboratory for Alignment , author=. 2021 , eprint=

2021
[43]

2020 , eprint=

The Curious Case of Neural Text Degeneration , author=. 2020 , eprint=

2020
[44]

2023 , eprint=

Faith and Fate: Limits of Transformers on Compositionality , author=. 2023 , eprint=

2023
[45]

2015 , eprint=

Distilling the Knowledge in a Neural Network , author=. 2015 , eprint=

2015
[46]

Maybank, and Dacheng Tao

Gou, Jianping and Yu, Baosheng and Maybank, Stephen J. and Tao, Dacheng , year=. Knowledge Distillation: A Survey , volume=. International Journal of Computer Vision , publisher=. doi:10.1007/s11263-021-01453-z , number=

work page doi:10.1007/s11263-021-01453-z
[47]

2023 , eprint=

The False Promise of Imitating Proprietary LLMs , author=. 2023 , eprint=

2023
[48]

2023 , eprint=

Large Language Models are Zero-Shot Reasoners , author=. 2023 , eprint=

2023
[49]

2022 , eprint=

Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? , author=. 2022 , eprint=

2022
[50]

2023 , eprint=

Mathematical Capabilities of ChatGPT , author=. 2023 , eprint=

2023
[51]

2020 , eprint=

Scaling Laws for Neural Language Models , author=. 2020 , eprint=

2020
[52]

2022 , eprint=

Training Compute-Optimal Large Language Models , author=. 2022 , eprint=

2022
[53]

2023 , eprint=

Scaling Relationship on Learning Mathematical Reasoning with Large Language Models , author=. 2023 , eprint=

2023
[54]

2024 , eprint=

Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models , author=. 2024 , eprint=

2024
[55]

Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=
[56]

2023 , eprint=

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , author=. 2023 , eprint=

2023
[57]

, author=

Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=
[58]

Distilling system 2 into system 1

Distilling system 2 into system 1 , author=. arXiv preprint arXiv:2407.06023 , year=

work page arXiv
[59]

Self-training elicits concise reasoning in large language models

Self-Training Elicits Concise Reasoning in Large Language Models , author=. arXiv preprint arXiv:2502.20122 , year=

work page arXiv
[60]

C3ot: Generating shorter chain-of- thought without compromising effectiveness

C3oT: Generating Shorter Chain-of-Thought without Compromising Effectiveness , author=. arXiv preprint arXiv:2412.11664 , year=

work page arXiv
[61]

Can language models learn to skip steps? arXiv preprint arXiv:2411.01855, 2024

Can language models learn to skip steps? , author=. arXiv preprint arXiv:2411.01855 , year=

work page arXiv
[62]

Tokenskip: Con- trollable chain-of-thought compression in llms

Tokenskip: Controllable chain-of-thought compression in llms , author=. arXiv preprint arXiv:2502.12067 , year=

work page arXiv
[63]

Cot-valve: Length-compressible chain-of-thought tuning

CoT-Valve: Length-Compressible Chain-of-Thought Tuning , author=. arXiv preprint arXiv:2502.09601 , year=

work page arXiv
[64]

Demystifying long chain-of-thought reasoning in llms

Demystifying Long Chain-of-Thought Reasoning in LLMs , author=. arXiv preprint arXiv:2502.03373 , year=

work page arXiv
[65]

Reasoning Language Models: A Blueprint,

Reasoning Language Models: A Blueprint , author=. arXiv preprint arXiv:2501.11223 , year=

work page arXiv
[66]

L1: Controlling how long a reasoning model thinks with reinforcement learning

L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning , author=. arXiv preprint arXiv:2503.04697 , year=

work page arXiv
[67]

Dast: Difficulty-adaptive slow-thinking for large reasoning models

DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models , author=. arXiv preprint arXiv:2503.04472 , year=

work page arXiv
[68]

Kimi k1.5: Scaling Reinforcement Learning with LLMs

Kimi k1. 5: Scaling reinforcement learning with llms , author=. arXiv preprint arXiv:2501.12599 , year=

work page internal anchor Pith review arXiv
[69]

Think smarter not harder: Adaptive reasoning with inference aware optimization.arXiv preprint arXiv:2501.17974, 2025

Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization , author=. arXiv preprint arXiv:2501.17974 , year=

work page arXiv
[70]

O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning

O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning , author=. arXiv preprint arXiv:2501.12570 , year=

work page arXiv
[71]

Training language models to reason efficiently

Training Language Models to Reason Efficiently , author=. arXiv preprint arXiv:2502.04463 , year=

work page arXiv
[72]

arXiv preprint arXiv:2502.03860 , year=

BOLT: Bootstrap Long Chain-of-Thought in Language Models without Distillation , author=. arXiv preprint arXiv:2502.03860 , year=

work page arXiv
[73]

When more is less: Understanding chain-of-thought length in llms

When More is Less: Understanding Chain-of-Thought Length in LLMs , author=. arXiv preprint arXiv:2502.07266 , year=

work page arXiv
[74]

Let’s think dot by dot: Hidden computa- tion in transformer language models

Let's think dot by dot: Hidden computation in transformer language models , author=. arXiv preprint arXiv:2404.15758 , year=

work page arXiv
[75]

From explicit cot to implicit cot: Learning to internalize cot step by step

From explicit cot to implicit cot: Learning to internalize cot step by step , author=. arXiv preprint arXiv:2405.14838 , year=

work page arXiv
[76]

Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach

Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach , author=. arXiv preprint arXiv:2502.05171 , year=

work page internal anchor Pith review arXiv
[77]

Training Large Language Models to Reason in a Continuous Latent Space

Training large language models to reason in a continuous latent space , author=. arXiv preprint arXiv:2412.06769 , year=

work page internal anchor Pith review arXiv
[78]

Compressed chain of thought: Efficient reasoning through dense representations

Compressed chain of thought: Efficient reasoning through dense representations , author=. arXiv preprint arXiv:2412.13171 , year=

work page arXiv
[79]

Efficient Reasoning with Hidden Thinking

Efficient Reasoning with Hidden Thinking , author=. arXiv preprint arXiv:2501.19201 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[80]

Advances in Neural Information Processing Systems , volume=

Simpo: Simple preference optimization with a reference-free reward , author=. Advances in Neural Information Processing Systems , volume=

Showing first 80 references.