RAG Safety, DeFi Workflows, Speech Retrospection, and KV Cache Optimization

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering alternative data integration for financial forecasting, safety benchmarks for RAG systems and DeFi workflow synthesis, inference-time backdoor detection, reinforcement learning with transition lookahead, speech LLM reasoning, external KV caching performance, and language-guided geometric optimization for CAD.

1. Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting

Authors: Jihoon Kwon, Lawrence Liu, Daekyung Park, Sumin Kim, Haverty Jack et al. | Categories: cs.AI Link: arxiv.org/abs/2609.11607

The paper investigates whether LLMs can forecast firm performance by integrating alternative data (consumer transactions, web traffic, prediction markets) with financial information via in-context learning. It proposes a two-agent framework that identifies which firms benefit from each data channel and then predicts revenue using firm- and channel-specific context. Experiments across four commercial alternative data channels show that adding alternative data in context improves forecasting accuracy over either source alone and outperforms standard baselines.

Takeaway: This work demonstrates a practical path for leveraging heterogeneous, sparse alternative data in finance without task-specific model retraining, a compelling template for applying LLM flexibility to real-world quantitative problems.

2. DeFiFlowBench: Benchmarking and Improving Safe Executability in Natural-Language DeFi Workflow Synthesis

Authors: Abhinav Rajeev Kumar, Harshit Arora, Varun Singh, Manikandan Nanjappan | Categories: cs.LG, cs.SE Link: arxiv.org/abs/2609.11504

DeFiFlowBench is introduced as a benchmark of 207 prompts for natural-language DeFi workflow synthesis that measures graph coverage, configuration completeness, and safety predicates, tested on a local EVM. Standard prompting approaches produce 14–19 unsafe executions per configuration, and quote-derived slippage bounds fail to prevent price impact. The proposed Koan-Safe framework combines intent parsing, replaceable generation, and structural repair with default safety parameters, achieving no unsafe executions on held-out prompts where baselines scored far worse.

Takeaway: A timely reminder that structurally valid generated code can still authorize costly trades, and that execution-based evaluation with explicit safety enforcement is essential for LLM-driven DeFi automation.

3. RAG-Safety-Bench: Reliable Evaluation of Retrieval-Augmented LLM Safety

Authors: Adithiyan Rajan Indira Saravanan, Kathleen C. Fraser | Categories: cs.CL, cs.IR Link: arxiv.org/abs/2609.11758

RAG-Safety-Bench isolates the safety impact of retrieval by controlling retriever quality across four conditions: non-RAG, oracle documents, related-but-answer-free documents, and random safe documents. Results across five open-source LLMs reveal an inverse relationship between benign and unsafe capability, show that baseline safety guardrails don’t guarantee downstream safety in RAG settings, and provide model-specific evidence that even benign retrieved documents can trigger unsafe generation.

Takeaway: As enterprises increasingly wire RAG into internal knowledge bases, this benchmark exposes a critical blind spot: retrieval safety is not inherited from model safety, and deployment teams need evaluation protocols beyond standard guardrail checks.

4. Language-Augmented Semantic Priors for B-Spline Surface Fitting

Authors: Yunzhong Lou, Yusheng Luo, Jiahao Li, Yu Song, Xiangdong Zhou | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2609.11708

LASP introduces a language-augmented framework that uses fine-tuned LLMs to infer structured, solver-usable B-spline priors from CAD procedural modeling histories. Operating as a semantic reasoning layer above existing geometric kernels, it translates modeling histories into textual descriptions of design intent, then predicts B-spline prior parameters through a two-stage training scheme. Experiments show that language-driven reasoning serves as a powerful inductive bias for geometric solving, improving semantic coherence over traditional ML approaches.

Takeaway: A novel bridge between the largely disconnected worlds of natural-language design intent and numerical geometry kernels, pointing toward CAD systems that understand not just shapes but the reasoning behind them.

5. Near-Optimal Reinforcement Learning with Multi-Step Transition Lookahead

Authors: Corentin Pla, Hugo Richard, Marc Abeille, Vianney Perchet | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2609.11807

The authors resolve two open questions about RL with multi-step transition lookahead: they prove exact planning remains NP-hard for every fixed rational discount factor in (0,1), and they introduce a randomized polynomial-time approximation scheme for every fixed lookahead depth. Extending to unknown transitions and stochastic rewards via optimism and variance-adaptive confidence bounds, their algorithm achieves cumulative regret matching classical tabular discounted RL up to logarithmic factors.

Takeaway: An elegant resolution of a hardness question that closes the gap between worst-case intractability and practical near-optimal learning, with regret guarantees that should interest both theorists and practitioners of lookahead-augmented RL.

6. RetroThinker: Enabling Retrospective Thinking in Speech LLMs

Authors: Yi-Jen Shih, Puyuan Peng, Abdelrahman Mohamed, David Harwath | Categories: eess.AS, cs.AI, cs.CL Link: arxiv.org/abs/2609.11864

RetroThinker is a multi-stage post-training framework that equips the Moshi speech LLM to self-verify and forward-correct chain-of-thought steps during streaming inference. It combines supervised fine-tuning on curated retrospective thinking data with length-based DPO to optimize reasoning that occurs concurrently with user speech. On GSM8K, RetroThinker achieves an 11% absolute accuracy gain at comparable latency compared to non-retrospective baselines, significantly improving the accuracy-latency trade-off.

Takeaway: Rather than accepting the usual accuracy-latency tradeoff in streaming speech reasoning, this work shows that a model can revise its own reasoning mid-stream, a promising direction for real-time voice assistants that need to think while listening.

7. Building py-kvcache: A Performance Characterization of External KV Caching for vLLM with NVMe SSDs

Authors: Joseph Kanichai, Tiziano De Matteis, Animesh Trivedi | Categories: cs.DC, cs.LG Link: arxiv.org/abs/2609.11744

This paper characterizes KV cache tradeoffs in vLLM across GPU, CPU, and NVMe tiers, finding that cache performance depends on transfer granularity, intermediate memory use, and scheduling timing rather than device bandwidth alone. The authors build py-kvcache, a vLLM KV Offload connector with asynchronous direct I/O and scheduler-aware preloading that starts disk reads while requests wait. At 80k tokens, py-kvcache loads 2.0x faster than LMCache, though on an H100 the average request falls below break-even, showing external caching is a setup-specific admission decision.

Takeaway: A refreshingly honest systems paper that doesn’t oversell external KV caching, offering concrete design principles and an open-source implementation while acknowledging when simply keeping prefixes in GPU memory wins.

8. SpecGuard: Inference-Time Backdoor Detection For Free

Authors: Rui Wen, Ahmed Salem, Andrew Paverd, Mark Russinovich, Zheng Li | Categories: cs.CR, cs.CL Link: arxiv.org/abs/2609.11799

SpecGuard repurposes speculative decoding as a zero-cost inference-time backdoor detector: when a backdoor triggers, the target model shifts toward attacker behavior while the clean draft model does not, causing draft-token acceptance rates to change measurably. The authors formalize when this signal appears and show that suppressing it requires weakening the backdoor itself. Across diverse backdoor types and model families, SpecGuard reliably detects triggered behavior, including stealthy cases that evade input-level filters, without extra generation cost.

Takeaway: A clever recycling of existing inference infrastructure, turning speculative decoding’s draft-verification gap into a free, always-on security monitor that stealthy attacks can’t easily evade.

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