Sharpen Without Search: On-Policy Distillation of Sequence-Level Power Distribution

On-policy power distillation trains models to generate sharpened answers directly, improving single-sample math reasoning by up to 27.3 points and outperforming multi-candidate sampling and reward-based methods.

Erfan Baghaei Potraghloo, Seyedarmin Azizi, Arya Fayyazi, Saeid Shokoufa and 3 more

Published Oct 5, 2026 · ▲ 4 on Hugging Face · Code ★ 1

Base Models Can Reason By Taking a Cue From Training Data

Fixing initial token cues in base models boosts reasoning to match RL performance, with effects traced to training data associations that can be causally edited.

Sophie L. Wang, Amil Dravid, Rulin Shao, Kevin Farhat and 2 more

Published Oct 5, 2026 · ▲ 12 on Hugging Face · Code ★ 2

What Matters for Latent Reasoning with Flow Matching

FLaRe uses flow matching for latent reasoning that is useful, diverse, explainable, refinable and efficient, reaching 97% of explicit chain-of-thought accuracy at 25% latency.

Yassine Ouali, Adrian Bulat, Georgios Tzimiropoulos

Published Oct 5, 2026 · ▲ 10 on Hugging Face

Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability

Efficient reasoning training differs in impact: faithfulness usually drops due to inconsistency, but monitorability remains robust.

Samuel Lewis-Lim, Xingwei Tan, Mario Sänger, Zhixue Zhao and 1 more

Published Oct 2, 2026 · 0 citations · ▲ 13 on Hugging Face

Language Models that Play Chess and Explain Their Moves

Queen, a 4B-parameter chess-language model, plays at grandmaster level and explains moves via cross-attention to a silent expert encoder and iterative Bellman-style explanation distillation, surpassing larger frontier models.

Adithya Bhaskar, Jeffrey Cheng, Danqi Chen

Published Oct 2, 2026 · 0 citations · ▲ 31 on Hugging Face · Code ★ 18

Hierarchical Continuous Diffusion Language Models

HC-DLM couples discrete token generation with a continuous latent trajectory via a unified variational denoising objective, outperforming diffusion baselines on Sudoku, Countdown, and language modeling.

Hui Ren, Zihan Li, Chang Liu, Huidong Liu and 1 more

Published Oct 1, 2026 · 0 citations · ▲ 89 on Hugging Face · Code ★ 57

Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent Adversaries

BDA treats multi-LLM council moves as observations of a per-agent reliability model to yield calibrated answer probabilities and robustly handle persistent adversaries without extra LLM calls.

Ionel Eduard Stan, Paolo Napoletano

Published Oct 1, 2026 · 0 citations

Does Learning Protein Folding Generalize to Broader Reasoning?

Post-training on protein-folding data via discrete answers and continuous geometry improves structure prediction and broad reasoning across ten benchmarks.

Yong Liu, Zhanpeng Shi, Yizhou Dang, Zhongyue Zhang and 3 more

Published Sep 30, 2026 · 0 citations · ▲ 120 on Hugging Face · Code ★ 28

Interpolated Policy Distillation: A Controllable Continuum Between Off-Policy and On-Policy Distillation

Interpolated Policy Distillation mixes student and teacher token distributions to balance trajectory quality and learnability, outperforming off-policy and on-policy distillation across reasoning benchmarks.

Youxu Shi, Yifan Sun, Dacheng Yin, Haomiao Tang and 4 more

Published Sep 29, 2026 · 0 citations

TeacherGRPO: Closing the Capacity Gap in Reasoning Distillation via Teacher Alignment

TeacherGRPO aligns teachers to student distributions via reinforcement learning to overcome reasoning distillation's Gap Curse and improves student performance.

Zhenyu Lei, Zihan Chen, Yaochen Zhu, Shangbin Feng and 4 more

Published Aug 20, 2026 · 0 citations

On the Geometry of On-Policy Distillation

On-policy distillation updates occupy a sparse, low-dimensional parameter subspace that is functionally sufficient and geometrically distinct from supervised fine-tuning and reinforcement learning.

Zhennan Shen, Yanshu Li, Qingyu Yin, Chak Tou Leong and 5 more

Published Jun 5, 2026 · 0 citations · ▲ 75 on Hugging Face

RelayLLM: Efficient Reasoning via Collaborative Decoding

RelayLLM enables small language models to dynamically invoke large models for critical reasoning tokens via collaborative decoding, reducing costs by 98.2% while achieving 49.52% accuracy.

Chengsong Huang, Tong Zheng, Langlin Huang, Jinyuan Li and 2 more

Published Jan 8, 2026 · 0 citations · ▲ 30 on Hugging Face · Code ★ 41

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

DeepSeek-V3.2 improves efficiency via sparse attention, scaled reinforcement learning matching GPT-5, and agentic synthesis, with a special variant surpassing GPT-5 and reaching gold-medal IMO and IOI levels.

DeepSeek-AI, Aixin Liu, Mei, Aoxue, Lin, Bangcai and 36 more

Published Dec 2, 2025 · 8 citations · ▲ 274 on Hugging Face

SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs

SoftCoT uses a fixed assistant and projection module to generate soft reasoning tokens that boost LLM reasoning via parameter-efficient fine-tuning.

Yaogeng Xu, Xu Guo, Zhiwei Zeng, Chunyan Miao

Published 2025 · 14 citations

Continuous Audio Thinking for Large Audio Language Models

Gyojin Han, DongJae Lee, Changho Choi, Jongsuk Kim and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Hypothesis generation and updating in large language models

Huadong Xiong

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Reasoning with Sampling: Cutting at Decision Points

Felix Zhou, Anay Mehrotra, Quanquan C Liu

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

A Solvable Model of Chain-of-Thought in In-Context Learning

Kaito Takanami, Cengiz Pehlevan

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Contractive Restoring Flows: Robust Reasoning Distillation via Orbital Stability

Dongqi Zuo, Yuanyuan Wang, Chuan Zhou, Haoxuan Li and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Programmatic Reasoning with Structural Schema: A Unified Framework for Multi-Table Inference

Jialin Chen, Brandon Mayer, Michael Galkin, Sami Abu-El-Haija and 2 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Tio: Language Models with Parallel Streams of Thoughts, Inputs and Outputs

Guinan Su, Yanwu Yang, Xueyan Li, Jonas Geiping

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

LLMs Keep Thinking When Told Not To

Dianqiao Lei, Kevin Qinghong Lin, Pan Lu, Philip Torr and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Renoise Consistency: Unlocking Efficient Self-Correction for Diffusion Large Language Models

JIHOON LEE, Yeongbin Seo, Jaehyung Kim

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Chain-of-Correction: Progress-Aware Policy Steering via Anchor-Grounded Predictive Reasoning

Xuening Zhang, Xiang Deng, Jiayi Lin, Qi Lv and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Reasoning Pathologies in Large Language Models: A Diagnostic Perspective

Rohan Surana, Junda Wu, Sheldon Yu, Gagan Mundada and 5 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

COMPASS: Composable Policy-Amortized Structured Search for LLM-Based Optimization Modeling

Nguyen Le Minh Hoang, Van D Cuong, Bui Trong Duc, Huynh Thi Thanh Binh

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

TARP: Trace-Anchored Regularization Prior for Retaining Stepwise Reasoning

Wei Li, Tao Feng, Hangjie Yuan, Aojun Lu and 1 more

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Evaluating Whether LLMs Can Reliably Connect the DOTs?

Eftekhar Hossain, John Salvador, Santu Karmaker

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

From Solver Trajectories to Teaching Trajectories: Cognition-Aligned Reasoning Distillation

Yashuo Luo, Qianren Mao, ZiqiQin, Junnan Liu and 5 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Geometric Instability of Hidden-State Trajectories Predicts Reasoning Failures in Large Language Models

Hoda Fakharzadehjahromy, Andreas C Bueff, Fredrik Heintz, Mattias Tiger and 3 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

When Form Changes but Logic Doesn’t: Building Logic-invariant LLMs through Structures

Xuyuan Liu, Xinshuai Dong, Elynn Chen, Yujun Yan

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

MemPilot: Learning Transferable Latent Memory Mechanisms for LLM Reasoning

Changlong Shi, LINHAO LUO, Shigeng Chen, Guibin Zhang and 3 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

SteadyThought: Mitigating LLM Under-Thinking via Thought-Level Preference Optimization

Zhenyue Peng, Lin Yang, Jie Wang, Xiaoqi Ni and 6 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Stability in Multi-Step Reasoning via Jacobian-based Error Accumulation Analysis

Dongyue Li, Alice Duan, Ziniu Zhang, Hongyang Zhang

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Large Language Models as Graph Computational Solvers via Topology-aware Residual Attention

Wei Zhuo, Siqiang Luo

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation

Taehoon Kim, Seunggeun Cho, Dongsu Han

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Failure Profiling and Reachable Trajectory Selection for Reasoning Distillation

yijun Zhu, Jianxin Wang, Chengchao Shen

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Large language models suffer from a curse of ambiguity

Nicolas Zucchet, Hyun Dong Lee, Scott Linderman

Paris Poster Session 4, Thu, Dec 10, 5:30 PM–7:30 PM, Paris Poster Hall · Published 2026

Attention-Based Soft Answer Sets

Wael AbdAlmageed

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

MinPath: Learning Efficient LLM Reasoning via Minimal Dependency Paths

Zhijing Yang, Shuming Hou, Guohui Xiao, Lemei Zhang and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Interleaved Latent Thinking and Adaptive Termination for Efficient Reasoning LLMs

Zhao Jin, Rong-Cheng Tu, Wenhao Sun, Qi Guo and 3 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Too Smart to Teach: The Articulability Ceiling in Language Models

Shervin Ghasemlou

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

Understanding the Out-of-Distribution Generalization of Chain-of-Thought Reasoning in LLMs

Nuojing Liang, Xiaotong Yuan, Pan Zhou

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue

Dingdong WANG, Shujie LIU, Jinyu Li, Yuxuan Hu and 5 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Uni-Synergy: Bridging Understanding and Generation for Personalized Reasoning

zijun shen, Ruichuan An, Sihan Yang, Ziyu Guo and 4 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Correct but Unselectable: The Hidden Interface Tax in Multi-Candidate Reasoning

Qianli Ma, Zhiqing Tang, Fanshuai Meng, Zhi Yao and 2 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Chain-of-Thought Oversight Should Not Treat Faithfulness as Monitorability

Sichao Li, Sai Ma, Xiyang Hu, Chudi Zhong and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Why Transformer-Based Language Models Need Explicit Mechanisms of Cognitive Control

Suketu Patel, Hongbin Wang, Jin Fan

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Position: Robust Reasoning Requires Internal Time, Not Scale Alone

Amar Gahir, Shreyank Gowda

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Preserving Exploration for LLM Reasoning via Mean Order-Statistic Alignment

Ruotian Peng, Yi Ren, Long-Fei Li, Yandong Wen

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Commit Then Explore: Reasoning Models Diversify, Not Converge

Ben Jenkins, Mihaela Cardei

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Mining Logic under Uncertainty: Probabilistic Soft Logic with Energy-Based Inference for Chain-of-Thought Verification

Jiang Yu, Jinlong Tian, Kewei Cheng, Yue He and 6 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Illusory Pattern Perception Drives Spurious Inference in Large Language Models

Peihua Mai, Zhuoyan Shao, Xinbao Qiao, Meng Zhang and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

CoTrek: Toward Scalable On-Policy Distillation for Long Chain-of-Thought Reasoning

Heng Zhang, Chengyu Zhou, Jiajun Wu, Estella Liu and 7 more

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Insight-Driven Search: A Framework for Multi-objective Automated Heuristic Design with Large Language Models

Shunyu Yao, Fei Liu, Ji Cheng, Mingxuan Yuan and 3 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Toward Efficient Reasoning of Large Language Models via Latent Concept-Pyramid Modeling

Sijia Chen, Ningxin Su

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Understanding Reasoning from Pretraining to Post-Training: Chess as a Controlled Testbed

Jingyan Shen, Ang Li, Salman Rahman, Yifan Sun and 3 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Anchoring Reasoning Distillation via Syntactic Constraints

Zehua Cheng, Wei Dai, Jiahao Sun

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth

Xutao Ma, Somayeh Sojoudi

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Time to Pay Attention! Understanding High Complexity Corpus Reasoning Tasks

Prasann Singhal, Amanda Bertsch, Jacob Steinhardt, Sewon Min

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Externalized CPDAG Summaries Improve LLM Causal Deduction

Structured Thinking externalizes typed CPDAG summaries before reasoning, raising LLM causal deduction F1 by up to 13.4 points on Corr2Cause.

Wentao Sun, João P Nogueira, Dominique Verchere, Mathieu Acher and 1 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Ask, Answer, and Detect: Role-Playing LLMs for Personality Detection with Question-Conditioned Mixture-of-Experts

ROME uses role-playing LLMs to generate questionnaire answers from posts, then routes them via mixture-of-experts to improve personality detection and mitigate label scarcity.

Yifan Lyu, Liang Zhang

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

M2A: Synergizing Mathematical and Agentic Reasoning in Large Language Models

M2A synergizes mathematical and agentic reasoning via parameter-space model merging, improving SWE-Bench Verified solved rates from 44.0% to 51.2% without retraining.

JunJian Wang, Xin Zhou, Qiran Xu, Kun Zhan

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Neurosymbolic Learning for Inference-Time Argumentation

ITA is a neurosymbolic framework that trains LLMs to generate scored arguments for ternary claim verification, yielding predictions faithfully computed from explicit argument structures.

Gabriel Freedman, Adam Dejl, Adam Gould, Mansi - and 3 more

Paris Poster Session 4, Thu, Dec 10, 5:30 PM–7:30 PM, Paris Poster Hall · Published 2026

Quantized Reasoning Models Think They Need to Think Longer, but They Do Not

Post-training quantization of reasoning models increases chain-of-thought length and overthinking errors without improving accuracy, yet penalizing overthinking markers reduces reasoning cost and fixes failures.

Sanae Lotfi, Polina Kirichenko, Steven Li, Zechun Liu

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face

An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models

LRMs show a large production-evaluation gap, scoring near 48% on reasoning evaluation versus near-perfect production due to answer confirmation bias.

Mingzhong Sun, Teresa Yeo, Armando Solar-Lezama, Tan Zhi-Xuan

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 6 on Hugging Face

Structural Rationale Distillation via Reasoning Space Compression

D-RPC distills reasoning via reusable reasoning paths to stabilize supervision, improving student performance with fewer tokens.

Jialin Yang, Jiankun Wang, Jiajun Wu, Henry Leung and 2 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

From Table to Cell: Attention for Better Reasoning with TABALIGN

TABALIGN improves multi-step table reasoning by pairing diffusion planners generating binary cell masks with attention verifiers, raising accuracy 15.76 points and accelerating execution 44.64%.

Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang, Chunhe Wang and 5 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

INFUSER: Influence-Guided Self-Evolution Improves Reasoning

INFUSER co-evolves a question generator and solver via influence-guided rewards, improving reasoning by over 20% on math benchmarks without curated data.

Siyu Chen, Miao Lu, Beining Wu, Heejune Sheen and 6 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Hint Tuning: Less Data Makes Better Reasoners

Hint Tuning calibrates reasoning depth by using an instruct model as a difficulty probe, cutting tokens by 24-66% with 1K samples.

Siqi Fan, Minghao Li, Xiaoqian Ma, Xiusheng Huang and 5 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Reasoning as Compression: Unifying Budget Forcing via the Conditional Information Bottleneck

Conditional Information Bottleneck frames reasoning as lossy compression with a semantic surprisal prior, improving LLM reasoning efficiency with minimal accuracy loss.

Fabio Valerio Massoli, Andrey Kuzmin, Arash Behboodi

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 23 on Hugging Face

Asking the Right Questions: Improving Reasoning with Generated Stepping Stones

ARQ introduces a question generator that produces transferable intermediate stepping stones, improving reasoning LLM performance via fine-tuning on synthetic data.

Hengyuan Hu, Tingchen Fu, Minqi Jiang, Alexander Miller and 2 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions

A framework samples coherent but cognitively unavailable "alien" research directions by maximizing idea coherence while minimizing existing community availability, broadening explored vocabularies 3.5-7x over LLM baselines.

Alejandro H. Artiles, Martin Weiss, Levin Brinkmann, Iyad Rahwan and 5 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Procedural Refinement by LLM-driven Algorithmic Debugging for ARC-AGI-2

ABPR couples LLMs with Prolog meta-interpreters to debug programs via proof-tree analysis, achieving up to 98.33% Pass@2 on ARC-AGI-2 and extending to relational reasoning benchmarks.

Yuning Qiu, Lin-Feng Zou, Jiong-Da Wang, Xue-Rong Yuan and 1 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models

LLMs exhibit confirmation bias by proposing confirming triples rather than falsifying hidden rules, reducing discovery rates, though prompting counterexample consideration improves success from 42% to 56%.

Ayush Rajesh Jhaveri, Anthony GX-Chen, Ilia Sucholutsky, Eunsol Choi

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling

Recon scores reasoning traces by action reconstruction fidelity to avoid post-hoc rationalization in user modeling, yielding up to 70% win rates over baselines across domains.

Alan Zhu, Mihran Miroyan, Carolyn Wang, Andrew Zhou and 3 more

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Lattice Deduction Transformers

Lattice Deduction Transformer approximates sound deduction via lattice-projected recurrent states, achieving near-perfect Sudoku and Maze accuracy with small parameters and abstention guarantees.

Liam Davis, Alberto Alfarano, Leopold Haller, Mark Santolucito

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision

Under outcome-only supervision, scaling training-time reasoning length improves OOD performance after ID saturation via stronger inductive biases and reduced shortcut reliance.

Yihao Xue, Allan Zhang, Jianhao Huang, Amit Sahai and 1 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Rethinking On-Policy Self-Distillation for Thinking Models

Privileged self-distillation degrades thinking models by suppressing reasoning forks and self-correction tokens, reducing long-rollout accuracy by up to 17%.

Simran Kaur, Narutatsu Ri, Yinghui He, Liam Fowl and 1 more

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 2 on Hugging Face

Trust, but Don’t Verify: Epistemic Blind Spots in LLM Source Evaluation

LLMs detect fabricated statistics in isolation but ignore numeric validity during multi-source synthesis, weighing sources by analytical register rather than accuracy.

Rohan N Pradhan, Steve Goley

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Soft Token Alignment for Cross-Lingual Reasoning

SOLAR aligns soft-token representations across languages during supervised fine-tuning to improve multilingual reasoning consistency, boosting accuracy up to 17.7 points with largest gains on low-resource languages.

Ivy He, Jungsoo Park, Wei "Coco" Xu, Alan Ritter

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

LACE: Lattice Attention for Cross-thread Exploration

LACE enables parallel reasoning paths to share insights and correct errors during inference via cross-thread attention, improving reasoning accuracy by over 7 points.

Yang LI, Zirui Zhang, Yang Liu, Chengzhi Mao

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

LARK: Learnability-Grounded Trajectory Selection for Efficient Reasoning Distillation

LARK selects reasoning trajectories by student learnability via a training-loss rate factor and chi-squared-regularized policy, improving reasoning distillation efficiency and generalization.

Tianrun Yu, Kaixiang Zhao, Chih-Chun Chen, Amanda Hughes and 4 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Learning from the Self-future: On-policy Self-distillation for dLLMs

d-OPSD applies on-policy self-distillation to diffusion LLMs via suffix conditioning and step-level supervision, cutting optimization steps by ~90% versus RLVR while outperforming baselines on reasoning benchmarks.

Yifu Luo, Zeyu Chen, Haoyu Wang, Xinhao Hu and 3 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 174 on Hugging Face · Code ★ 18

ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence

ChainSpace structures spatial reasoning as state-preserving multi-round chains to expose hidden failures and improve data-efficient training.

Xiaohan Zhang, Feng Gu, Xudong Rao, Xuhao Pan and 3 more

Paris Poster Session 2, Wed, Dec 9, 5:00 PM–7:00 PM, Paris Poster Hall · Published 2026

On the Token Value Inequality in Efficient Reasoning

Token value inequality in reasoning traces enables identifying core versus redundant tokens via log probabilities, yielding 76% token reduction with preserved accuracy via selective compression.

Runjia Zeng, Hang Hua, Yiyang Liu, Zhiqiang Tao and 4 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

KL for a KL: On-Policy Distillation with Control Variate Baseline

vOPD casts on-policy distillation as RL and stabilizes it via a closed-form per-token reverse KL baseline that reduces gradient variance without extra inference cost.

Minjae Oh, Sangjun Song, Gyubin Choi, Yunho Choi and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Diversity Combining for Multi-Path LLM Reasoning

Multi-path LLM reasoning is modeled as diversity combining, showing path correlation limits majority-vote gains and that adaptive sampling retains near-peak accuracy.

Guangsheng Yu, Litianyi Zhang, Qin Wang, Xu Wang and 4 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · Code

A Measure-Theoretic Analysis of Reasoning: Structural Generalization and Approximation Limits

Reasoning is formalized via optimal transport to bound transformers' OOD generalization by architectural Lipschitz continuity and approximation limits, proving depth is needed for backtracking and shift-invariant attention reduces risk.

Yuyang Zhang, Yifu Zhang, Xuehai Zhou, Xiaoyin Chen

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Geometric Latent Reasoning Induces Shorter Generations in LLMs

Geometric Latent Reasoning predicts continuous embedding-space reasoning paths via iterative direction updates, yielding substantially shorter correct generations without length penalties.

Shashi Kumar, Yacouba Kaloga, Petr Motlicek, Ina Kodrasi and 1 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning

DiscoLoop combines discrete embeddings and continuous hidden states in looping transformers to fix representational misalignment, enabling near-perfect multi-hop reasoning with faster training and stronger pretraining performance.

Hengyu Fu, Tianyu Guo, Zixuan Wang, Hanlin Zhu and 4 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Interleaved Head Attention

Interleaved Head Attention mixes attention heads via pseudo-heads to enable cross-head reasoning, cutting parameters on synthetic tasks and improving retrieval and math benchmarks over standard multi-head attention.

Sai Surya Duvvuri, Chanakya Ekbote, Rachit Bansal, Rishabh Tiwari and 5 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 2 on Hugging Face

OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping

OS-Pruner formulates chain-of-thought pruning as optimal stopping to learn dynamic termination, cutting generation length 20-60% with minimal accuracy loss.

Mohammed Ehab, Aymane El Gadarri, Vivek Farias, Adam Jozefiak and 1 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning

Extracting search trees from LLM reasoning traces reveals myopic planning where performance depends on breadth rather than depth, unlike human planning.

Sixing Chen, Ji-An Li, Saner Cakir, Sinan Akcali and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

MUX: Continuous Reasoning via Multiplexed Tokens

MUX compresses reasoning into continuous multiplexed tokens via lossless superposition, accelerating reasoning and outperforming latent baselines across 32 settings.

Ayhan Suleymanzade, Halil Alperen Gozeten, Michael Bronstein, Ismail Ilkan Ceylan and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves

Depth-recurrent attention mixtures (Dreamer) combine sequence, depth, and sparse expert attention to scale latent reasoning efficiently, requiring 2, 8x fewer training tokens than matched baselines while improving expert diversity.

Jonas Knupp, Jan Metzen, Jeremias Bohn, Georg Groh and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face

When Reasoning Meets Its Laws

This paper proposes Laws of Reasoning (LoRe), a framework formalizing reasoning compute and accuracy laws, plus LoRe-Bench showing models lack compositionality; enforcing compute-law compositionality via finetuning improves reasoning performance.

Junyu Zhang, Yifan Sun, Tianang Leng, Jingyan Shen and 3 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 62 on Hugging Face · Code ★ 38

LeAct: Learning to Reason from Expert Actions

LeAct recovers expert reasoning chains from actions alone to train reasoning models; it reaches near-optimal expert performance across games and robotics while improving on direct imitation.

Ziran Yang, Chengshuai Shi, Raj Ghugare, Benjamin Eysenbach and 2 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026