Financial LLMs, DeFi Safety, RAG Evaluation, and KV Cache Optimization

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering alternative data forecasting with LLMs, DeFi workflow safety, RAG safety evaluation, language-guided CAD surface fitting, reinforcement learning with transition lookahead, retrospective reasoning in speech LLMs, external KV caching for vLLM, and inference-time backdoor detection.

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

Alternative data sources like consumer transactions and web traffic offer timely financial signals, but their limited historical coverage and heterogeneous nature make them hard to integrate into conventional forecasting. The authors propose a two-agent LLM framework that identifies which firms each data channel informs, then predicts revenue using firm- and channel-specific context. Experiments across four commercial data channels show combining alternative data with financial information outperforms either source alone and beats standard baselines.

Takeaway: A practical demonstration that in-context learning can flexibly absorb messy, sparse alternative data—an approach likely to resonate with quant researchers frustrated by traditional feature engineering pipelines.

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

Structurally valid DeFi workflows can still authorize costly trades, so the authors introduce DeFiFlowBench, a 207-prompt benchmark measuring graph coverage, configuration completeness, and safety predicates with execution on a local EVM. Standard prompting yields 14-19 unsafe executions per configuration, and quotes-derived slippage bounds fail to prevent price impact. Their Koan-Safe system combines intent parsing and structural repair with default safety parameters, scoring 0.67 on the static safety proxy versus 0.33 for the best baseline.

Takeaway: A sobering look at how easily LLM-generated financial workflows can go wrong, with an execution-based evaluation methodology that other agentic-finance efforts should adopt.

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

Retrieval-augmented generation can improve reliability but may inadvertently degrade safety when prompted for harmful content. RAG-Safety-Bench isolates the effects of retriever quality by testing four conditions—non-RAG, oracle harmful documents, related-but-answerless documents, and random safe documents—across five open-source LLMs. Results show an inverse relationship between benign and unsafe capability, and that baseline guardrails offer no downstream safety guarantee in RAG settings; even benign retrieved documents can trigger unsafe generation.

Takeaway: Essential reading for anyone deploying RAG over corporate knowledge bases—the finding that safety alignment doesn’t transfer cleanly to retrieval-enabled systems is a critical blind spot.

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

Traditional CAD kernels rely on heuristic initialization for B-spline surface fitting and largely ignore the procedural semantics embedded in modeling histories. LASP introduces a semantic reasoning layer above existing solvers: it translates modeling histories into textual descriptions of design intent, then uses a fine-tuned LLM to predict structured B-spline prior parameters. Experiments show language-driven reasoning acts as a powerful inductive bias, producing more semantically coherent fitting results than conventional machine learning schemes.

Takeaway: A clever bridge between LLM reasoning and classical geometric solvers—if it generalizes, this paradigm of language-guided optimization could reshape how CAD systems interpret design intent.

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

Prior work showed optimal planning with multi-step transition lookahead is NP-hard only for discount factors arbitrarily close to one, leaving open whether hardness holds generally and whether near-optimal planning is tractable. The authors prove exact planning remains NP-hard for every fixed rational discount factor, yet introduce a randomized polynomial-time approximation scheme for any fixed lookahead depth. Extending to unknown transitions and stochastic rewards via optimism and variance-adaptive bounds, they achieve regret matching classical tabular discounted RL up to logarithmic factors.

Takeaway: A satisfying resolution of a theoretical gap—hardness and tractability coexist here, and the regret bound suggests lookahead can be exploited practically without paying a statistical price.

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

SpeechLLMs offer lower latency than cascaded ASR-plus-LM pipelines but lag on complex reasoning, and real-time interaction imposes strict latency constraints that create an accuracy-latency tradeoff. RetroThinker is a multi-stage post-training framework that equips the Moshi model to self-verify and forward-correct chain-of-thought steps during inference, combining supervised fine-tuning on retrospective data with length-based DPO. On GSM8K, it achieves an 11% absolute accuracy gain at comparable latency over non-retrospective baselines.

Takeaway: Streamed self-correction of reasoning traces is a promising direction for real-time voice agents—the 11-point gain without added latency suggests retrospective thinking may become standard in speech LLM post-training.

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

Prefix caching can cut time-to-first-token for long-context LLM requests, but for short prefixes or fast GPUs, recomputation may beat loading from external caches. The authors characterize this tradeoff across GPU, CPU, and NVMe tiers in vLLM, finding performance depends on transfer granularity, intermediate memory use, and scheduling rather than raw bandwidth alone. Their py-kvcache connector with asynchronous direct I/O and scheduler-aware preloading is 2.0x faster than LMCache at 80k tokens, though on H100s the break-even point shifts and external caching becomes a setup-specific admission decision.

Takeaway: Valuable systems-level guidance for anyone operating vLLM at scale—the nuance that external KV caching isn’t universally beneficial, and the scheduler-aware preloading trick, are immediately actionable.

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

Deployed LLMs may carry hidden backdoors, but existing runtime detectors either assume trigger forms or require extra model computation that hurts latency. SpecGuard repurposes speculative decoding’s draft-token acceptance rate as a free signal: when a backdoor fires, the target model shifts toward attacker behavior while the clean draft model fails to predict it, altering acceptance rates. Across diverse backdoor types and model families, SpecGuard reliably detects triggered behavior—including stealthy cases invisible to input filters—with zero added generation cost.

Takeaway: Elegant repurposing of an existing optimization as a security monitor—turning speculative decoding into an always-on backdoor detector for free is the kind of systems insight the field needs more of.

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