Agents for Games and Driving, Fact-Checking Evidence, and Open Speech Models

Today’s selection of 8 noteworthy AI/ML papers from arXiv covers major advances across several frontiers: from LLM-driven game environment generation for reinforcement learning and agentic frameworks for autonomous driving testing, to foundational models for speech processing, physically-grounded dexterous robotics, and large-scale agentic retrieval benchmarks.

1. PlayTrain: An Efficient Reinforcement Learning Framework for LLM-Generated Adaptable JavaScript Games

Authors: Ryan Truong, Lance Ying, Samuel J. Gershman, Kazuki Irie | Categories: cs.LG Link: arxiv.org/abs/2609.09059

PlayTrain is a novel RL framework that leverages LLMs to generate JavaScript games from minimal human prompts and runs them efficiently in a standard gym environment. It supports end-to-end pixel-based agent training at over 1M agent-decisions per second on a single GPU, enabling cloning of Atari and ProcGen games as well as creation of modified versions with novel test sets or game dynamics. The framework reimagines the VGE development pipeline by reducing the entire process to a single, LLM-modifiable JS file.

Takeaway: This paper tackles the long-standing bottleneck of hand-coded RL environments by making LLM-generated games first-class citizens, potentially democratizing and accelerating RL research through rapid, adaptable environment prototyping.

2. PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners in Autonomous Driving

Authors: Yuan Gao, Sebastian Müller, Mattia Piccinini, Marc Kaufeld, Yuchen Zhang et al. | Categories: cs.AI, cs.CL, cs.RO Link: arxiv.org/abs/2609.08965

PlannerForge is the first unified LLM-agent framework that covers the entire scenario-based testing pipeline for Autonomous Driving Systems, from generation and selection to ADS enhancement and benchmarking. Chaining the modules end-to-end retains 83% of seed queries for commercial LLMs and 78% for open-source models, while outperforming existing baselines on generation, selection, and editing tasks. Cost-tuning at N=400 lifts planner success rates from 50.4% to 70.2% and cuts collisions from 19.0% to 8.4%, without domain-specific fine-tuning.

Takeaway: By unifying the fragmented modular pipeline for ADS testing, PlannerForge makes open-source models like Qwen3.6:35B viable competitors to commercial APIs, potentially lowering the barrier to robust scenario-based validation in autonomous driving.

3. Evaluating and Improving Evidence-Grounded Fact-Checking in LLMs via Multi-Round Evidence Ablation

Authors: Xingyu Deng, Mingzi Cao, Nikolaos Aletras, Xi Wang, Mark Stevenson | Categories: cs.CL, cs.AI, cs.IR Link: arxiv.org/abs/2609.08943

This paper introduces Fact-Ablated Evaluation (FAE), a framework that iteratively ablates cited evidence to reveal that off-the-shelf LLMs rely more on parametric knowledge than on provided evidence when fact-checking. To address this, the authors propose REAL (Rigorous Evidence Ablation Learning), a training method that uses counterfactual evidence supervision to promote true evidence dependency. Experiments across four fact-checking datasets show models trained with REAL achieve superior evidence-dependent capabilities compared to standard fine-tuning.

Takeaway: This work exposes a hidden failure mode in LLM-based fact-checkers — where strong performance can coexist with weak evidence grounding — and offers a training recipe that aligns predictions more faithfully with verifiable sources.

4. TriCCOT: Tri-part Convolutional Conformal Transformer for Onboard Space Object Detection

Authors: Adrien Dorise, Marjorie Bellizzi, Julia Cohen, Stéphane May | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2609.08659

TriCCOT is a tri-part architecture for deployable onboard space object detection that combines a convolutional region proposal network, a conformal prediction stage for distribution-free coverage guarantees, and Aper-GATES, a hardware-friendly attention-based classifier that reformulates self-attention with convolutional projections and gating operations. Experiments on DIOR and VDVRaw datasets show competitive detection performance and improved robustness to spatial blur and signal-dependent noise versus FPGA-compatible baselines. The model is fully deployed on a Xilinx Versal VCK190 FPGA without modifying the underlying DPU architecture.

Takeaway: TriCCOT bridges the gap between hardware-efficient CNNs and global-reasoning transformers, showing that unified CNN-Transformer inference is feasible on existing accelerators for spaceborne applications.

5. DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination

Authors: Yankai Fu, Ning Chen, Junkai Zhao, Heng Zhang, Guocai Yao et al. | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2609.09119

DeCAL is a physically-grounded dexterous vision-language-action model built on a Mixture-of-Transformers architecture that unifies understanding, imagination, and action generation for contact-rich manipulation. It introduces Adaptive Visuo-Tactile Fusion, which dynamically gates tactile interactions via contact-awareness, and Visuo-Tactile Latent Co-Imagination to jointly model visual and tactile dynamics. DeCAL achieves a 71% average success rate and 83.4% progress success rate, consistently outperforming state-of-the-art baselines while demonstrating strong generalization to unseen scenarios.

Takeaway: By explicitly modeling contact dynamics and adaptively integrating tactile signals rather than relying on homogeneous fusion, this work pushes toward physically-grounded policies — a key step toward robust dexterous manipulation in the wild.

6. AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing

Authors: Ziyang Ma, Zhikang Niu, Wenming Tu, Tianrui Wang, Ruiqi Yan et al. | Categories: cs.SD, cs.CL, cs.MM Link: arxiv.org/abs/2609.08936

AuK is an open-source foundational model unifying speech generation and editing through natural-language instructions and audio context, trained on approximately 3.03 billion instruction-audio instances across five task families. Its architecture combines a multimodal LLM for semantic conditioning, a joint VAE for acoustic conditioning, and a hybrid rectified-flow Transformer with dual-stream MMDiT blocks. The distilled AuK-Flash model performs 4-step inference without classifier-free guidance, achieving a 4.5x wall-clock speedup while maintaining leading performance on zero-shot generation and general instruction-guided editing.

Takeaway: This work represents a significant open-source contribution to the speech domain, offering both full and distilled models that rival commercial systems across a broad — and previously fragmented — range of speech tasks.

7. Evidence-Grounded Retrieval for Investigation Hunt Lead Generation from CTI Reports

Authors: Akash Prakash, Boubakr Nour, Makan Pourzandi, Chadi Assi, Mourad Debbabi | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2609.08790

AHLERT automatically extracts actionable, environment-aware hunt leads from threat reports, combining a hybrid retriever (dense vector search plus multi-hop KG traversal seeded with MITRE ATT&CK) with ontology-grounded RAG that constrains leads to the defender’s own assets and controls. On public CTI reports for well-known APTs, hybrid evidence retrieval with ontology grounding raises mean F1 by ~2x (0.44 to 0.85) over a flat-RAG baseline, achieving an effectiveness score of ~86.95%. The framework is LLM-agnostic and emits structured, directly actionable leads.

Takeaway: This paper addresses a practical gap in threat hunting — converting unstructured CTI reports into environment-specific hypotheses — with a system that is show to work across multiple proprietary and open-weight LLMs.

8. Q2D-Web: A Large-Scale Benchmark for Retrieval in Agentic RAG Systems

Authors: Maximilian Schall, Sedigheh Eslami, Markus Krimmel, Antoine Chaffin, Louis Milliken et al. | Categories: cs.IR, cs.CL Link: arxiv.org/abs/2609.08887

Q2D-Web is a large-scale agentic retrieval benchmark comprising a 190M-document web corpus and 70k agentic search queries in ten languages, reformulated from real-world user queries in production systems. It provides three sets of fixed relevance judgments (agent citations, production rankings, and a combined set with LLM-based judgments) to reduce false negatives. Benchmarking 13 retrievers, the authors find model ranking is largely insensitive to the judgment set but diverges substantially across domains, languages, and query types; they also show that retaining a third of the corpus preserves the full-corpus ranking while raising absolute Recall@1000 by only 3-7 points.

Takeaway: Q2D-Web fills a critical evaluation gap for agents that reformulate queries — the interface between human intent and first-stage retrievers that standard search benchmarks miss.

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