October LLM, SLM, and Agent Research: Local Models, Tool Selection, and Memory
October research offers local models worth testing and separates small-model tool selection, answer verification, and confidence calibration. Enterprise evaluation also needs memory, authorization, and workload tests.
Nemotron 3.5 Lightning has an official GB10 deployment recipe worth comparing with existing local models; 0.8B OvisOCR3 merits a separate document evaluation. Start SLMs on verifiable subtasks, while the platform controls handoff, stale memory, authorization, and shared resources.
As of October 30, 2026, we reviewed 12 papers first submitted that month and revisited the earlier Sleep research. This article draws on arXiv manuscripts and official model cards. Dates refer to first submission, with revisions noted separately. Results belong to the authors' experiments; X Cube has not reproduced them on its DGX environment. Our aim is to identify useful comparisons and the conditions those comparisons must preserve.
The October 28 Nemotron 3.5 report introduces two local-model candidates. Super is a multimodal model with 120B total and 12B active parameters; Lightning is a text model with 30B total and 3B active parameters. Lightning's official model card includes a DGX Spark and GB10 recipe. That makes evaluation more practical, but capacity planning must still include all weights, caches, workspace, and concurrent requests. Active parameter count describes computation rather than total model-loading memory. [1][2]
Nemotron 3.5 Bolt, first submitted October 15, explores compression and low-bit quantization with 74B total and 9B active parameters. Its 24GB estimate has specific conditions, including batch size one, precision, and context, and excludes the vision encoder and projector. The paper's performance platform also differs from this project's Linux service. Multimodal capability, text-serving memory estimates, and actual serving cost therefore need separate checks. [3][4]
Specialized small models deserve their own evaluation. The October 20 OvisOCR3 report builds on Qwen3.5-0.8B for OCR, document parsing, information extraction, and cross-page tables. Official weights and inference examples are available. Traditional Chinese forms, scanned documents, and dense tables still require a local test set. Field accuracy, table structure, missing text, and incomplete output are closer to document-workflow requirements than a general model ranking. [5][6]
Training Small Language Models for Transferable Tool Selection, submitted October 22, trains Qwen3 models at 1.7B, 4B, and 8B using a Plan, Retrieve, Select workflow. Evaluation requires the selected tool set to match the annotated set and tests transfer to other tool corpora and retrievers. This supports tool selection as a trainable subtask. Adoption still requires checking the training assets and data, then separately measuring argument correctness and complete workflow success. [7]
Handoff needs to be evaluated alongside verification. The October 6 STEPGATE paper decides at each step whether to execute locally, verify, or use a stronger model. Its stronger backend is a 7B model from the same family, and its task setting is bounded. The October 28 Blind Spots paper separates solving tasks, identifying successful trajectories, and predicting success. More sampling does not eliminate verification and calibration gaps. Routing needs calibrated signals and independent acceptance checks rather than self-reported confidence alone. [8][9]
Harness-Aware Distillation, submitted October 2, studies the division of work between a model and its execution software. When software supplies state, available actions, and failure records, distillation can focus on how the student uses that information. Experiments use text interaction environments; the authors leave extensions to real tool-calling and software agents for future work. For enterprise platforms, this suggests retaining inspectable rules and state, collecting replayable trajectories, and then testing which capabilities a small model needs to learn. [10]
Agent memory first needs to track whether a fact remains valid. The October 26 paper When Retrieval Is Not Enough shows that, on its synthetic single-hop update chains, two Qwen models can be distracted by obsolete facts even when the latest fact is retrieved. More evidence does not necessarily improve the answer. Evaluation should distinguish retrieval of the latest version, inclusion of older versions, and the version ultimately used. Provenance, validity periods, and replacement relationships need to travel with memory. [11]
The October 27 Proactive Memory paper asks when an agent should intervene. It manages recent events, source-validated temporal trajectories, and behavioral regularities in separate layers, then chooses intervention or abstention. Experiments focus on low-risk Sports, Home, and Code settings; independent held-out evaluation and cross-model transfer remain open. Enterprise evaluation also needs to measure whether notifications help, how much they interrupt, and whether abstention is appropriate. [12]
Tenant boundaries and authorization require checks across steps. CoSA, submitted October 26, studies how individually acceptable skills can violate policy when their artifacts or decisions are reused for another purpose. It motivates checking information flow and authorization before externalization, persistent mutation, or execution. A tool belonging to an organization, a user being able to see it, and a particular action being authorized are distinct decisions that need distinct test cases. [13]
Correct business-process order also requires more than tool descriptions. The October 27 paper Grounding LLM Agents in Business Processes combines OData schemas, value constraints, and BPMN process relationships to guide service invocation. Its mock validator shares a schema with the knowledge graph, however, and an external corpus without a process layer shows no gain. A practical experiment starts with one real workflow, compares tool descriptions, structured schemas, and process constraints, and checks independent business outcomes. [14]
Concurrency should be judged by delivered work. AOS-Bench, first submitted October 23 and revised October 27, compares dynamic workloads sharing CPU and memory and finds a gap between isolated task performance and multi-task delivery. Its tasks use FIFO and lack deadlines, priorities, and dependencies. Those conditions do not establish a universal concurrency setting. Platforms need measurements of waiting, resource use, cancellation, recovery, and tenant fairness before choosing capacity policies. [15]
Sleep needs a separate date and mechanism check. Language Models Need Sleep was first submitted June 2 and revised July 10, rather than first appearing in October. It studies distillation, replay, and reinforcement learning to consolidate short-term knowledge into model parameters. Enterprises exploring this route must define training-data provenance, tenant-specific or shared parameter updates, and rollback. Task scheduling, external-memory updates, and model learning need separate management. [16]
For X Cube, useful next experiments compare existing Ling and Qwen models with Lightning, assess OvisOCR3 on Traditional Chinese documents, and test SLM tool selection or step-level handoff. Measure task success, tool-argument correctness, latency, memory, and cost alongside obsolete-fact and cross-tenant or cross-purpose negative cases. These are proposed experiments; promotion of a model still depends on local results.
Source
- Nemotron 3.5 Super and Lightning: technical report
- Nemotron 3.5 Lightning: official model card and GB10 recipe
- Nemotron 3.5 Bolt: compression, inference, and memory estimates
- Nemotron 3.5 Bolt: official model card
- OvisOCR3: technical report
- OvisOCR3: official model card and inference examples
- Training Small Language Models for Transferable Tool Selection
- Blind Spots: Where Small Language Models Fall Behind Frontier Models on Terminal Tasks
- Do I Need the Cloud? Uncertainty-Aware Step-Level Handoff for Small Language Model Agents
- Harness-Aware Distillation for Small Language Model Agents
- When Retrieval Is Not Enough: Obsolete Facts in Agent Memory
- Proactive Memory for Event-Driven LLM Agents
- CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems
- Grounding LLM Agents in Business Processes
- AOS-Bench: Can LLM Agents Manage Dynamic Workloads? (v2)
- Language Models Need Sleep (submitted in June, revised in July)


