Tag
A practitioner weighs three designs for adding persistent knowledge-base spreadsheets to a chat agent's existing tools, leaning toward a dedicated analyze_knowledge_spreadsheet tool with clear routing descriptions and a routing test, while asking the community for missed approaches.
An in-depth breakdown of Lecture 5, "Agent Tools in Pictures," from Columbia University's Agentic Engineering course — 221 minutes covering the essence of tools, the complete call chain for a single invocation, five ways to package tools (functions / CLI / Skills / MCP / code execution), and security plus engineering judgment. The core thesis: "The model writes the order, the program does the cooking — prompts are not the lock, code is."
ToolRACER introduces a synthetic data generation pipeline that emulates user, assistant, and tool interactions to produce adversarial multi-turn conversation trajectories for training robust task-oriented agents, along with ToolRACERBench, a 5.6K-trajectory benchmark spanning six domains with ~66% failure-prone scenarios. Models trained on it show improved accuracy on τ²-bench and ACEBench.
Mosaic is a constrained decoding framework for diffusion language models that significantly improves performance on function calling benchmarks like JevBench and is compatible with vLLM, SGLang and Hugging Face, supporting models such as DiffusionGemma and LLaDA2.
The post argues that as an agent's tool set grows, direct tool calling degrades, and the more robust fix is letting the model write code in a sandbox to call tools as functions — reducing schema clutter in context and lowering wrong-call rates — at the cost of sandbox security concerns.
This research paper extends FunctionGemma 270M for practical on-device Android workflows by introducing a synthetic dataset and fine-tuning the model, achieving improved accuracy for function calling while balancing performance and coverage.
Google AI developers demonstrated how they used Gemini 3.8 Live Extended Thinking to build a coding tutor that assists with debugging by analyzing the screen and referencing the p5.js library.
This paper introduces a closed-world resolution method to combat tool hallucination in LLM agents, offering a taxonomy and benchmark for measuring and addressing fabricated tool calls.
Cactus Needle 3 is a small, sliceable foundation model for automation tasks that runs on-device, achieving performance comparable to larger models on function calling and structured extraction.
Jev by typesafe AI enables function calling to execute in real-time as users type, understanding intent and arguments without waiting for Enter.
This paper introduces a carbon-aware routing framework for function-calling LLMs in edge-cloud systems, reducing operational carbon emissions by an average of 4× while maintaining cloud-level accuracy.
A research paper presents a method for training a smaller self-hosted LLM using separate GRPO experts merged via SLERP, which outperforms a larger baseline on enterprise tasks and serves half of platform traffic at lower cost.
Liquid AI releases DSpark draft models for their LFM series, incorporating speculative decoding to achieve up to 4x decode speedup on device while maintaining output quality.
This paper investigates the impact of constrained decoding on tool-call abstention, decomposing the effects of grammar constraints into stop and emission components, and evaluates performance on small open-weight models across languages.
LiquidAI announces LFM2.5-VL-3B, an efficient vision-language model for edge hardware with improved screen understanding, grounding, multi-image input, and function calling, trained with 4x more vision data and post-training via SFT and RL.
InclusionAI (Ant Group's lab) launched Ling 3.0 Tiny, a closed API model with 1.3B activated params out of 7.9B total, 256K context, native function calling, and a thinking/instant mode, aimed at multi-turn agent tool loops. The article questions the efficiency curve and notes no independent evals or open weights.
Data Turnstile is an open-source framework for generating high-quality synthetic function-calling training data from API specifications. Fine-tuning small language models with this data significantly improves their tool-use performance, closing the gap with much larger models.
SAAG proposes a cascaded diagnostic framework for evaluating LLM agent function calling by decomposing evaluation into registry conformance, structural completeness, and argument grounding stages, enabling interpretable diagnostics and iterative self-repair. Experiments with sub-4B models show improved argument precision and reduced value hallucination compared to single-pass evaluation.
GnLOLot releases GGUF quantizations of the MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking model, a 1B parameter thinking model fine-tuned on Fable 5 data with improved tool/function calling compared to V1, designed for local deployment via llama.cpp and compatible runtimes.
This article introduces how to upgrade single-turn RAG to Agentic RAG, by allowing the LLM to autonomously decide on multiple retrievals and tool calls to solve multi-step reasoning for complex problems. It provides code examples and implementation ideas based on Qwen3.7-Max.