AI & Automation Briefing - June 26, 2026
JSON Schema Constraints Silently Kill Tool Calls in Open-Weight Models
A new paper from HuggingFace researchers identifies a failure mode they call "Tool Suppression," which occurs when JSON Schema structured output constraints and tool calling are enabled together on open-weight models. Across seven models tested, enforcing JSON Schema constraints via grammar-based token masking significantly reduced how often models actually invoked tools, while still producing schema-compliant output. The researchers term the underlying decoding-level mechanism "Constraint Tax." For builders running self-hosted models in agentic workflows, this means a model can silently stop firing tool calls without throwing any errors, producing clean JSON that contains no actions. The practical implication is that structured output constraints and tool calling capability cannot be assumed to coexist reliably in open-weight models without explicit testing and architectural accommodation.
LLM Agents Don't Remember Their Plans, They Just Re-Read Them
A new paper from Mehta and Datta tests whether LLM agents carry plan information forward as persistent internal state or simply depend on the plan text staying in context. Using a diagnostic method called replay pairing, they measured hidden-state similarity on Llama-3.1-70B and found that plan signal spikes immediately after planning, then drops 4.1x within a single action-observation step. When plan text was evicted from context in an ALFWorld stress test, task success fell by 34.7 percentage points. The practical implication for long-running agentic workflows is direct: any context compression, summarization, or truncation step that removes early plan content will silently degrade agent performance. Designing reliable multi-step automation requires treating plan text as protected context, not as information the model has already internalized.
Research Identifies Why Tool-Calling Agents Collapse and How Supervisory Signals Stabilize Them
A new paper from researchers at the Chinese Academy of Sciences pinpoints a specific failure mode in reinforcement learning (RL) trained tool-use agents: probability spikes in control tokens disrupt structured tool-invocation sequences, causing abrupt performance collapse. The underlying tool-use capability survives intact, but gets blocked by format breakdown. The researchers tested multiple supervisory approaches including off-policy supervision, hint-based guidance, and erroneous example supervision, under both synchronous and interleaved training schemes. Interleaving supervised fine-tuning (SFT) with RL produced the most stable training, though it showed weaker generalization when output formats or content shifted out of distribution. For anyone building or explaining multi-step AI agents, this research provides a concrete, evidence-based account of why chained tool-use behaviors are unreliable and what training interventions actually address the problem. Code is available at https://github.com/hypasd-art/Tool-RL-Box.