KV Cache Quantization, Robot Agent Benchmarks, and Knowledge Editing in LLMs

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering KV cache quantization for looped Transformers, adaptive evidence routing for RAG, decoupled knowledge editing via conditional memory, a robot-use benchmark for multimodal agents, label-free dataset pruning, industrial UI code generation, global point set alignment, and intent-aligned text-to-music generation.

1. ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals

Authors: Heejun Kim, Junyoung Lee, SangLyul Cho, Dongsu Han, Insu Han et al. | Categories: cs.LG Link: arxiv.org/abs/2610.10381

Looped Transformers gain computational depth by reusing shared blocks, but their KV cache memory scales with the number of loops, throttling batch size and throughput. ResidualQuant exploits the observation that KV states across loops are highly similar, using final-loop KV states as a reference and encoding remaining loops as low-precision residuals with least-square scaling, rotations, and loop-wise mixed precision. It retains near-BF16 accuracy while cutting theoretical KV storage by 80.7%, boosting decode throughput up to 2.73x and peak throughput up to 4.15x on an RTX 5090.

Takeaway: A clever, architecture-aware take on KV quantization that achieves accurate INT2 storage and major real-hardware speedups — a rare combination of memory savings and measured throughput gains.

2. RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing

Authors: Yilun Hao, Krishna Sayana, Isabella Ye, James S Ren, Sukhdeep Sodhi et al. | Categories: cs.AI Link: arxiv.org/abs/2610.10507

Conventional RAG assumes the needed evidence already exists in some source item, but many tasks require deriving evidence through filtering, aggregation, or computation across items. RECAST formulates evidence construction as a sequential decision process: a lightweight RouterLM selects and formulates operations (or custom operations for a frozen CompilerLM to compile into code) until evidence suffices, then passes it to a frozen AnswerLM. Trained with SFT and GRPO, it reaches a 75.6% mean success rate across six benchmark families, beating the strongest large-model baseline by 15.9%.

Takeaway: Reframes retrieval as an active computation-and-routing problem, and the margin over large baselines plus held-out generalization suggests the “derive, don’t just retrieve” framing is paying off.

3. EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

Authors: Hongru Cai, Ran Wei, Wenjie Wang, Chengfa Wu, Ning Song et al. | Categories: cs.CL Link: arxiv.org/abs/2610.10533

Conditional memory architectures like DeepSeek Engram can decouple factual storage from computation, but different expressions of a fact activate different n-gram embeddings, and shared-embedding updates can corrupt unrelated knowledge. EngramEdit computes target memory representations for multiple expressions of an updated fact, then jointly updates shared n-gram embeddings to match, penalizing changes to frequently reused embeddings. It achieves near-perfect edit success, generalizes across unseen expressions and multi-hop reasoning (nearly 3x the strongest baseline’s CoT accuracy), and largely preserves unrelated knowledge and general capabilities.

Takeaway: Turns conditional memory into a genuinely editable knowledge interface — a promising alternative to finetuning-based or locate-and-edit approaches for keeping LLM facts current.

4. RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

Authors: Zhiqin Yang, Chenxin Li, Xiaomeng Hu, Yibin Liu, Weidong Huang et al. | Categories: cs.RO, cs.LG Link: arxiv.org/abs/2610.10409

RobotWorld is a simulation testbed of 84 tasks spanning manipulation, mobile manipulation, locomotion, driving, and aerial control, with explicit interaction budgets and executable success checks. Analyzing task outcomes alongside execution traces, the authors find agents can build sophisticated perception-and-control workflows yet fail to compose them reliably — losing object state tracking, failing to correct ineffective actions, or mistaking unfinished tasks for completion. Failure profiles differ across models, with Astra stronger on spatial and contact-constrained goals and Opus 5.5 stronger on balance and timed interactions.

Takeaway: A rigorous, trace-level account of where digital agent competence does and doesn’t transfer to physical execution — valuable for anyone building embodied agents.

5. Dataset Pruning from First Principles: A Label-Free Linear Programming Approach

Authors: Rodrigo Schuller, Francisco Ganacim | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2610.10347

Rather than assuming nearby embeddings share properties, this work derives selection criteria by reformulating unbiased subset selection as variance minimization, characterizing unbiased selectors as a high-dimensional polytope where expected sampling variance is a linear objective. Because standard linear programming is impractical at that dimension, the authors build an efficient vertex walk over closed-form pairwise variance expressions — requiring no labels and no model training during selection. Across CIFAR-10, MNIST, and CelebA it matches or exceeds uniform sampling at every budget, with the same framework also reducing stochastic-gradient variance in mini-batches.

Takeaway: A principled, assumption-light foundation for dataset pruning that doubles as a mini-batch diversification trick — appealing for its theoretical clarity and zero-training-cost selection.

6. TaoD2C-Bench: Benchmarking MLLMs for Industrial UI Code Generation Beyond Visual Fidelity

Authors: Chengwei Shi, Yunnong Chen, Tingting Zhou, Qiang Lu, Shiyu Yue et al. | Categories: cs.SE, cs.AI Link: arxiv.org/abs/2610.10374

Industrial design-to-code requires MLLMs to combine UI images with messy layer metadata and realize layouts under target-library constraints — capabilities that go well beyond visual recognition. TaoD2C-Bench is built on 2,861 production designs from 17 commercial platforms with 97,652 expert annotations across Component, Group, Alignment, and Position, distinguishing required constraints from permitted implementation choices. Evaluating eight MLLMs across end-to-end code generation, requirement inference, and requirement realization reveals substantial gaps and shows that visual reconstruction ability does not imply constraint-satisfying code.

Takeaway: A realistic, expert-annotated benchmark that exposes the fidelity-versus-correctness gap in UI code generation — a useful reality check for multimodal coding claims.

7. Rubix: Global Correspondence-Free Point Set Alignment through Assignment Geometry

Authors: Subhransu S. Bhattacharjee, Dylan Campbell, Rahul Shome | Categories: cs.CV, cs.CG, cs.LG, cs.RO, math.OC Link: arxiv.org/abs/2610.10408

Procrustes-Wasserstein alignment jointly estimates matching and rotation without correspondences, but alternating minimization can stall at suboptimal solutions. Rubix solves the equally weighted planar problem globally under squared Euclidean loss by exploiting the permutation polygon: each matching defines a complex correlation, supporting vertices give optimal matchings at fixed rotations, and the farthest vertex gives the global alignment. The authors prove the sharp n(n−1) vertex bound (answering Rote’s open problem) and recover the polygon in O(n⁵) operations, achieving reference-accurate MPEG-7 matching in 12 ms on average — 50x faster than a rotation grid — with gains on 3D scans, shape retrieval, and noisy crystal classification.

Takeaway: A rare combination of a resolved open problem in computational geometry and a practically faster, globally optimal alignment method for vision and robotics.

8. MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent

Authors: Zekai Liu, Zhilin Wang, Xuzheng He, Yu Cheng, Yang Yang | Categories: cs.SD, cs.AI Link: arxiv.org/abs/2610.10355

Single text-audio relevance scores can mask failures in specific requirements like instrumentation, structure, rhythm, or mood progression, especially for underspecified prompts. The authors formulate text-to-music intent alignment as satisfying a per-request rubric of independently verifiable items covering explicit and implied intent, instantiated as MuRA-Bench with expert-curated real platform requests. MIRA is a test-time agent that grounds intent into rubrics, then searches prompt revisions for a black-box generator under a bounded budget using trajectory-aware tree search and rubric feedback. It improves intent alignment enough for an open-source backend to rival commercial systems like Suno and Mureka.

Takeaway: Rubric-based, diagnostic evaluation plus test-time prompt search is a smart way to close the intent gap in generative music without touching model weights.

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