Towards In-Parameter Memory Augmentation for Large Language Models

This survey organizes in-parameter memory augmentation for LLMs by parameter placement and acquisition time to enable reusable parametric knowledge at deployment.

Haoyu Huang, Zhongwei Xie, Jiaxin Bai, Yisen Gao and 5 more

Published Oct 6, 2026 · ▲ 5 on Hugging Face · Code ★ 1

Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation

Self-generated feedback in long-horizon test-time training causes weight updates that improve synthetic text but degrade real-text prediction, and settlement on independent evidence prevents this failure.

Cheng Luo, Bing Li, Bernard Ghanem

Published Oct 4, 2026 · ▲ 19 on Hugging Face · Code ★ 1

ALoDLM: Adaptively Looped Diffusion Language Models

ALoDLM applies token-adaptive latent recurrence to diffusion language models, allocating computation by difficulty to close the quality gap with autoregressive models at 1.7B and 8B scales.

Liancheng Fang, Zhuowei Li, Youngeun Kim, Tianchen Zhao and 9 more

Published Oct 3, 2026 · ▲ 55 on Hugging Face · Code ★ 4

ReSolve: Reusing Candidate Reasoning through Selective Generative Moderation

ReSolve reuses candidate reasoning via selective generative moderation to boost math accuracy and cut token use versus voting and self-consistency.

Bangji Yang, Jiajun Fan, MA Hongba, Xi Zhu and 5 more

Published Oct 1, 2026 · 0 citations

Learning to Predict Distributions over Weight Updates for Test-Time Adaptation

Query-conditioned hypernetworks predict distributions over LoRA weight updates from input queries, enabling test-time scaling via sampled adapted models that outperform deterministic and token-sampling baselines.

Azal Ahmad Khan, Keshav Ramji, Tahira Naseem, Ali Anwar and 1 more

Published Oct 1, 2026 · 0 citations

LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

AutoTTS automatically discovers test-time scaling strategies via environment-driven controller synthesis, improving LLM reasoning accuracy-cost tradeoffs over manual baselines at minimal cost.

Tong Zheng, Haolin Liu, Chengsong Huang, Huiwen Bao and 9 more

Published May 8, 2026 · 0 citations · ▲ 70 on Hugging Face · Code ★ 176

TTCS: Test-Time Curriculum Synthesis for Self-Evolving

TTCS uses a co-evolving question synthesizer and reasoning solver to build test-time curricula that stabilize self-updates and improve reasoning on hard math benchmarks.

Chengyi Yang, Zhiyi Xiang, Yunbo Tang, Zongpei Teng and 4 more

Published Jan 30, 2026 · 0 citations · ▲ 35 on Hugging Face · Code ★ 53

When One LLM Drools, Multi-LLM Collaboration Rules

Multi-LLM collaboration outperforms single LLM reasoning on tasks where individual models fail, demonstrating collective rule over solo drooling.

Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang and 9 more

Published 2026 · 1 citation

Efficient Test-Time Scaling via Self-Calibration

Self-Calibration distills self-consistency confidence into LLMs for reliable single-pass estimation, enabling confidence-based early stopping that improves MathQA accuracy to 83.6 with 16 samples.

Chengsong Huang, Langlin Huang, Leng, Jixuan, Jiacheng Liu and 1 more

Published Feb 25, 2025 · 1 citation · ▲ 15 on Hugging Face · Code ★ 22

Mean-Field Parallel Decoding for Discrete Diffusion Language Models

Tamim Zoabi, Ameen A Ali, Liran Ringel, Lior Wolf

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

Catch Your Breath: Adaptive Computation for Self-Paced Sequence Production

Alexandre Galashov, Matt Jones, Nan Rosemary Ke, Yuan Cao and 2 more

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

Early Signals, Strong Decisions: Prefix-Guided Sampling for Parallel Test-Time Scaling

Jie Ren, Jonathan S Rosenfeld, Neil Thompson

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

Test-time Scaling for Diffusion Language Models with Frequency-Aware Remasking

Bowen Zuo, Yue Yu, Dongruo Zhou, Yinglun Zhu

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

Relaxation-Aligned State Control for Test-Time Scaling in Generative Combinatorial Optimization

Bohao Li, Ying Li, Pei He, Yangming Guo

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

Confidence-Calibrated Inference Expansion for Evaluator-Guided Test-Time Reasoning

Weida Liang, Kenji Kawaguchi

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

A Model of Diverse Sampling from Language Models

Manuel Prada-Corral, Yahya Emara, Timothy O'Donnell, Ryan Cotterell and 1 more

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

Training Quality Determines Efficiency Boundaries in Test-Time Reasoning

MD Azizul Hakim

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

Sparse Internal Control of Language Models

xizhe zhang

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

Test-Time Prompt-Agnostic Decomposition

Junze Wang, Lei Fan, Dezheng Zhang, Donglin Di and 3 more

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

When Does Structure Help? Statistical Tradeoffs for Structured Reverse Processes in Diffusion Large Language Models

Ruofeng Yang, Jingyuan Liu, Shuai Li

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

Denoising Time Matters: Diverse Generation in Diffusion Language Models

jingxuan wu, Zhenglin Wan, Yuzhe YANG, Yiqiao Huang and 4 more

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

Specialists Hold, Generalists Discount: Asymmetric Equilibrium in LLM Routing Auctions

Xinyu Hou, Yang Lu, Rabimba Karanjai, Pei-Chi Pan and 3 more

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

When LLM Routers Overpay: Strong-Model Over-Selection under Loose Budgets

Guannan Lai, Long Chen, Han-Jia Ye

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

Latent Refinement Decoding: Enhancing Diffusion Language Models by Refining Belief States

Qinglin Zhu, Yizhen Yao, Runcong Zhao, Yanzheng Xiang and 7 more

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

Prefill-Guided Trace Allocation for Sample-Efficient Test-Time Scaling

Zhi Yao, Zhiqing Tang, Hanshuai Cui, Qianli Ma and 2 more

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

What should post-training optimize? A test-time scaling law perspective

Muheng Li, Jian Qian, Wenlong Mou

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

Achieve Latency-Efficient Temporal-Coding Spiking LLMs via Discretization-Aware Conversion

Jinjie Fang, Tianxing Man, Xiao Du, Chengxun Jin and 3 more

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

HiLoRA: Adaptive Hierarchical LoRA Routing for Training-Free Domain Generalization

Ziyi Han, Huanyu Wang, Zeyu Zhang, Xiangxiang Dai and 2 more

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

LLM Routing Through the Lens of Recommendation: A Roadmap for Efficient AI Orchestration

Yuchen Zhou, Boyu Wang, Di Wu

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

Geometric Prompt-Trajectory Planning for Test-Time Scaling

Zhengqi Pei, Anran Zhang, Qingming Huang, Shuhui Wang

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

Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery

Samidha Verma, Ankit Anand, Sayan Ranu

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

Gumbo: Gumbel Optimized High-Temperature Speculative Sampling

Jonah Yi, Dan Fu, Yu-Xiang Wang

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

SIGMA: A Sigmoid-Gated Sampler for Test-Time Scaling in Diffusion Language Models

Ziwen Zhang, Weiyu Chen, Yuyan Zhou, Yichen Zhu and 1 more

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

Training Agent to Scale Inference-Time Reasoning

Yunzhe Qi, Sirui Chen, Jiaru Zou, Yanjun Zhao and 2 more

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

Exact Regular-Constrained Sampling for Variable-Order Markov Generation

Francois Pachet

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

Not All Low-Confidence Tokens Are Equal: Calibrated Confidence for Efficient Test-Time Reasoning

Tangyu Jiang, Haodi Wang, Yuanbing Zhu, Xiaojiang Du and 2 more

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

StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning

Yang Ouyang, Jung-Eun Kim

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

Selected-Tail Reliability in Verifier-Guided Best-of-$N$ Inference

Teresa Zhang

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

Attention Drift: What Auto-Regressive Speculative Decoding Models Learn

Doğaç Eldenk, Payal Mohapatra, Yigitcan Comlek, Kaan Oktay and 2 more

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

Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification

ADAP adaptively increases response sampling and verification to find verified positives with near-optimal expected cost under monotonic rewards.

Shaddin Dughmi, Mahdi Haghifam, Yusuf H Kalayci

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face · Code ★ 2

STDec: Spatio-Temporal Stability Guided Decoding for dLLMs

STDec improves dLLM decoding speed via spatial and temporal stability-guided adaptive thresholds without training, achieving up to 14.17x speedup.

Yuzhe Chen, Jiale Cao, Xuyang Liu, Jin Xie and 2 more

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

Self-Rewarding Sequential Monte Carlo for Masked Diffusion Language Models

Self-rewarding sequential Monte Carlo improves masked diffusion language model sampling via trajectory-level confidence weights across parallel particles, boosting quality without training.

Ziwei Luo, Ziqi Jin, Lei Wang, Lidong Bing and 1 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 5 on Hugging Face · Code ★ 16

Enabling approximate joint sampling in diffusion LMs

A lightweight sampler layer on frozen diffusion LMs approximates joint token sampling, yielding MAUVE 0.87 versus 0.31 when unmasking four tokens per step.

Parikshit Bansal, Sujay Sanghavi

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

On the Overscaling Curse of Parallel Thinking: System Efficacy Contradicts Sample Efficiency

Parallel LLM reasoning wastes compute via global budgets; sample-specific predictions via LanBo and PreAda improve efficiency without sacrificing accuracy.

Yiming Wang, Zhuosheng Zhang, Rui Wang

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

Confidence-Based Decoding is Provably Efficient for Diffusion Language Models

Confidence-based decoding achieves ε-accurate diffusion language model sampling in Õ(H(X₀)/ε) iterations by adaptively unmasking tokens until cumulative entropy exceeds a threshold.

Changxiao Cai, Gen Li

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

Attention-Based Sampler for Diffusion Language Models

Attention-based sampling orders diffusion language model tokens by attention-matrix column sums to maximize likelihood and improve generation quality with greater parallelism.

Yuyan Zhou, Kai Syun Hou, Weiyu Chen, James Kwok

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

Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization

Joint Consistency frames test-time aggregation as energy minimization using pairwise interactions and evaluation signals, outperforming existing voting methods across reasoning benchmarks.

Yunzhen Yao, Hongye Wang, Yahong Wang, Michael Gastpar and 2 more

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

PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding

PSD accelerates diffusion LLM inference via adaptive parallel unmasking and multi-depth speculative drafts with hierarchical verification, achieving up to 5.5x tokens per pass with near-greedy accuracy.

Shengyin Sun, Yiming Li, Renxi Liu, Xinqi Li and 6 more

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

State of Thought Enables Endogenous Reasoning

SoT enables endogenous LLM reasoning via internal dynamics-geometric states and a lightweight controller, improving accuracy by up to 2.51x while reducing tokens by 62.6% and latency by 44.6% versus external reasoning methods.

Zhiren Gong, Yikun Hou, Zihao Zeng, Ming Xiao and 2 more

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

DMax: Aggressive Parallel Decoding for dLLMs

DMax improves diffusion language model parallel decoding via progressive self-refinement and on-policy uniform training, boosting tokens per forward pass by over 2.5x while preserving accuracy.

Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu and 1 more

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

Social Choice Foundations for Simulation-Augmented Generation

SAGE formalizes efficient inference-time viewpoint simulation via metric proportional justified representation, proving small simulated pools and dynamic routing preserve approximate proportional representation for contentious queries.

Sonja Kraiczy, Smitha Milli, Ratip Emin Berker, Avinandan Bose and 5 more

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

Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping

Learned stochastic stopping reduces out-of-distribution variance in looped transformers by decoupling loop count from sequence length during training. It improves accuracy-stability trade-offs across algorithmic tasks, though it can stabilize suboptimal computation.

Hsun-Yu Kuo, El Mahdi Chayti, Patrik Reizinger, Wieland Brendel and 1 more

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

Provable Test-Time Scaling for Beam Search in LLM Reasoning

Modified confidence-filtered beam search achieves near-linear token-level coverage scaling versus quadratic for vanilla beam search, with polynomial horizon dependence that outperforms exponentially scaling Best-of-N methods.

Qijia He, Yu Huang, Yuan Cheng, Yuxin Chen and 1 more

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

Don't Always Pick the Highest-Performing Model: An Information Theoretic View of LLM Ensemble Selection

Formulating ensemble selection as mutual-information maximization reveals an information-theoretic error floor from model correlation and yields a greedy algorithm that outperforms baselines under fixed query budgets.

Yigit Turkmen, Baturalp Buyukates, Melih Bastopcu

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

Consilience for Verifier-Free Test-Time Scaling

Confidence-based verifier-free test-time scaling fails on complex tasks because high initial confidence signals no exploration; consilience selects rollouts by requiring low early but high final confidence, improving reasoning and coding.

Lecheng Kong, Like Hui, Haitao Mao, Luke Huan

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 2 on Hugging Face

Predicting and improving test-time scaling laws via reward tail-guided search

Tail-guided reward estimation predicts LLM test-time scaling laws and guides SLG Search to dynamically allocate compute, achieving vanishing regret and polynomial compute savings over best-of-N.

Muheng Li, Jian Qian, Wenlong Mou

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

RheoSampling: Resolving the One-Hot Dilemma in Stochastic Dynamic-Tree Speculative Decoding

RheoSampling decouples tree construction and token verification via proxy probabilities to enable lossless stochastic dynamic-tree speculative decoding with higher acceptance rates and speedups.

Qiao Hu, Yepeng Weng, Bo Zhang, Takehisa Yairi

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

Ensembling Language Models with Sequential Monte Carlo

A byte-level sequential Monte Carlo algorithm samples from composed language model ensembles, outperforming naive probability averaging across structured generation tasks.

Robin Chan, Tianyu Liu, Samuel Kiegeland, Clemente Pasti and 4 more

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

Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference

This paper models parallel inference-time reasoning via particle filtering, deriving non-asymptotic guarantees and fundamental limits for sequential Monte Carlo with process reward models.

Noah Golowich, Fan Chen, Dhruv Rohatgi, Raghav Singhal and 3 more

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

DAWN: Dependency-Aware Fast Inference for Diffusion LLMs

DAWN extracts token dependency graphs to select reliable unmasking positions, accelerating diffusion LLM inference by 1.80-8.06x with negligible quality loss.

Lizhuo Luo, Zhuoran Shi, Jiajun Luo, Zhi Wang and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face · Code ★ 12