Efficient Reasoning, Divergent Token Uncertainty, Long-Term Memory, and Multimodal Continual Tuning

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering training-free reasoning efficiency via spectral null-space composition, uncertainty quantification through divergent token counting, auditable long-term memory retrieval, multimodal continual instruction tuning, statistical guarantees for deep anomaly detection, in-context dense retrieval, autonomous robot policy improvement, and residual-enhanced emotional speech generation.

1. $S^3$: Spectral Null-Space Swap Makes Reasoning Models Efficient

Authors: Hongbo Ma, Sansheng Cao, Jiajun Fan, Bangji Yang, Ge Liu | Categories: cs.LG, cs.AI, cs.CL, cs.CV Link: arxiv.org/abs/2609.37976

The reasoning capability of thinking models is found to reside in the weight component within the null space of a projection defined by the corresponding non-thinking model’s dominant singular directions, and removing this subspace component largely improves reasoning efficiency without hurting accuracy. The authors propose Spectral Null-Space Swap ($S^3$), a training-free composition of paired non-thinking and thinking checkpoints that keeps the non-thinking model within its dominant subspace while taking the thinking checkpoint outside it. Evaluated across 2B–30B dense and MoE architectures in 28 environments spanning math, multimodal, and audio reasoning, $S^3$ reduces inference token overhead by an average of 27.4% while improving task accuracy by 1.0 percentage point.

Takeaway: A rare training-free method that identifies a principled spectral decomposition of what makes reasoning models verbose, yielding simultaneous efficiency and accuracy gains—this reframes thinking-mode post-training as a subspace alignment problem rather than a distillation one.

2. Probability is Not Enough: Exploring and Counting Divergent Tokens for Reasoning Uncertainty Quantification in LLMs

Authors: Feiyang Li, Shengjing Liu, Qi Zhan, Sijie Cheng, Weiqing Wang et al. | Categories: cs.AI Link: arxiv.org/abs/2609.38070

The authors introduce Divergent Token Confidence (DTC), a framework that estimates model confidence by counting tokens where two models strongly disagree during decoding, identified via Jensen-Shannon divergence between next-token distributions along the same reasoning trajectory. The count of divergent tokens is almost negatively associated with answer accuracy, providing a simple and effective uncertainty signal without training or affecting generation. DTC achieves an average expected calibration error of 13.0% in white-box settings, compared to 32.7%–42.4% for standard full-sequence confidence methods, and improves calibration in black-box settings as well.

Takeaway: A surprisingly simple signal—disagreement counts between paired models—substantially outperforms probability-based confidence estimation, suggesting that the field’s fixation on token probabilities has been a calibration dead end.

3. Auditable Long-Term Memory: A Deterministic Retrieval Chain Measured at 479/475 of 500 on LongMemEval-S

Authors: Christopher J. Chanhnourack | Categories: cs.CL, cs.AI, cs.IR Link: arxiv.org/abs/2609.38021

This work evaluates an auditable long-term memory system on LongMemEval-S whose retrieval chain uses hybrid candidate retrieval, cross-encoder reranking, coverage-first packet compilation, and deterministic reasoning scaffolds, with an LLM used only as a replaceable final reader. Two 500-question passes score 479/500 and 475/500, with the chain placing all gold sessions in the candidate pool for 468/470 answerable questions. The author carefully notes that all components were developed on the same 500 questions with no held-out evaluation, judge re-scoring flips three verdicts on byte-identical answers, and headline reader requests were not fully retained.

Takeaway: The paper’s real contribution may be its unusually candid auditability report—explicitly cataloging where the evaluation falls short of establishing superiority—making it a valuable template for honest benchmarking in an era of inflated memory-system claims.

4. ReCAP: Retrieval-Guided Capability Reuse for Multimodal Continual Instruction Tuning

Authors: Tao Hu, Zhinuo Zhou, Xialiang Tong, De-Chuan Zhan, Da-Wei Zhou | Categories: cs.CV, cs.LG Link: arxiv.org/abs/2609.37889

ReCAP is a retrieval-guided framework that leverages external knowledge to guide capability reuse during multimodal continual instruction tuning, incrementally building a knowledge base of domain, reasoning, and format knowledge from current-stage training data. For each instruction, retrieved domain knowledge guides generation while retrieved reasoning knowledge selects and orders capability modules into an instance-specific capability path. To enable stable cross-stage reuse, the framework introduces adaptive subspace recycling, parameterizing reusable modules with shared bases and stage-specific cores to protect historically important directions while recycling residual capacity.

Takeaway: By externalizing knowledge retrieval and composing capability paths at inference time, ReCAP offers a compelling alternative to the parameter-constraint orthodoxy in continual learning—especially promising for multimodal settings where reasoning patterns genuinely transfer across tasks.

5. Post-Anomaly Detection Inference for Deep SVDD

Authors: Cao Le Cong Thanh, Dang Quang Vinh, Vo Nguyen Le Duy | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2609.37935

The authors propose PADI (Post-Anomaly Detection Inference), a framework that equips a trained and frozen Deep SVDD detector with statistically valid inference using the Selective Inference framework, performing inference conditional on a test instance being identified as anomalous. PADI derives valid selective p-values that quantify the statistical significance of detected anomalies and theoretically establishes false positive rate control at a user-specified significance level. The framework is extended to Deep Semi-Supervised Anomaly Detection (Deep SAD), with experiments on synthetic and real-world benchmarks demonstrating proper FPR control and superior true positive rates.

Takeaway: Deep anomaly detection has long lacked rigorous statistical guarantees for individual decisions; PADI bridges this gap by porting selective inference into the deep SVDD setting, a meaningful step for safety-critical deployments.

6. Effective Dense Retrieval using Only In-Context Examples

Authors: Nour Jedidi, Abdul Basit Ali, Hang Li, Jimmy Lin | Categories: cs.IR, cs.CL Link: arxiv.org/abs/2609.38099

The authors introduce RICE (Representations from In-Context Examples), a training-free approach that extracts high-quality dense representations from decoder-only LLMs by conditioning the model on examples that provide a shared context for query and document encoding. RICE embeddings substantially improve the accuracy of prompt-based LLM embeddings, establishing a simple method to build LLM-based dense retrievers without any retriever training.

Takeaway: If in-context examples can close much of the gap to trained dense retrievers, this raises pointed questions about how much of retriever training is really learning retrieval versus learning a format—worth testing at larger scales.

7. Skill-Space Shooting for Autonomous Robot Policy Improvement

Authors: Zihang Rui, Renhao Wang, Haoxu Huang, Yang Gao | Categories: cs.RO, cs.AI, cs.LG Link: arxiv.org/abs/2609.38178

Skill-space shooting uses foundation model guidance to explore corrections through reusable skills—familiar short behaviors that recur across tasks and can be reasoned about from a scene—and turns successful trials into policy improvement without requiring human demonstration of each correction. Real-world experiments show repeated improvement in policies acting autonomously, while skills can be shared across tasks to reduce the teaching needed for new tasks.

Takeaway: The insight that corrective behaviors are themselves reusable skills—and can serve as supervision—offers a practical path to self-improving robot policies that avoids the sample inefficiency of pure reinforcement learning from scratch.

8. EmoRES-TTS: Residual-Enhanced Vector Steering for Emotional Speech Generation

Authors: Kuan-Po Huang, Haohe Liu, Puyuan Peng, Haibin Wu, Zhaoheng Ni et al. | Categories: cs.SD, cs.CL, eess.AS Link: arxiv.org/abs/2609.38157

The authors discover that an emotion steering vector in TTS can be decomposed into a shared component that moves speech away from neutral expression and a residual component that directs generation toward the requested emotion, and propose EmoRES, a training-free method that controls both components without retraining the backbone. On IEMOCAP, EmoRES outperforms conventional vector steering across all four objective emotion metrics on IndexTTS-2 and CosyVoice2, improving rank correlation by 26.13 and 12.97 percentage points and emotion hit rate by 12.95 and 6.92 points. Human evaluation shows up to 35.0% relative improvement in listener identification of the requested emotion and up to 17.3% improvement in fidelity.

Takeaway: Decomposing emotion vectors into shared and residual components is a clean, interpretable fix to a known weakness of global-strength vector steering, and the training-free nature makes it immediately applicable to existing frozen TTS backbones.

This content was generated with AI assistance. Paper information sourced from arXiv.