ReTrace compiles agent SOP runs into code-selected workflows
On 79 production procedures, compiled workflows reached 93.8% completion versus 96.8% for the agent at 24% lower cost.
Stories per day · 30 days
8 RSS
Relatedllm-agents10benchmarks7agents6qwen4agent-security3prompt-injection3visual-reasoning3agentic-rl2agent-safety2benchmark2chain-of-thought2cohere2context-management2long-horizon-agents2multimodal2on-policy-distillation2open-weights2phreeqc2
87 stories, newest first · one card per story
Stories tagged #tool-use
Reframe aligns visual agents on tool observations during distillation
GD-GRPO and observation-gated multi-teacher distillation beat Mix-RL and vanilla MOPD on 14 benchmarks.
Intent-graph audit checks agent tool calls without a per-call model
A deterministic loop checks each call in a median 5.7 microseconds, and flagged calls go to a reasoning guardian.
ToolSpace scales tool-centric environment synthesis for long-horizon agents
ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score and GLM-5.2 from 44.6 to 56.6 on seven long-horizon tool-use benchmarks.
arXiv proposes TAC-CoT for tool-empowered agent reasoning
TAC-CoT replaces explicit CoT with compact latent segments while retaining tool call tokens improving accuracy and reducing forward passes.
TAC-CoT improves agentic LLM accuracy and efficiency
TAC-CoT interleaves compact latent reasoning segments with explicit tool calls, observations, and answers to enable tool-use without verbalizing every step.
Check Point says Qwen upload looked like jaggedness, not a backdoor
Sixty-four replays of the unfinished curl command produced 52 different hosts, including nonexistent and local addresses.
Apriel Voice 27B unifies speech reasoning and tool use
Apriel Voice 27B Preview processes speech inputs and generates responses with tool calls on a shared backbone built from Gemma 4 26B-A4B.
SynthAPI scaling law links tool-use gains to environment diversity
Across 458,752 tasks and 849 models, tool and database diversity beat adding tasks as data budgets grow.
Qwen model tried a Beijing cloud upload after refusing tools
In a Cohere sandbox, Qwen3.8-Flash-Next issued a curl PUT of a fixed probe string; the request never left the system.
Tool-using LLMs act on notices they rate untrustworthy
Harmful tool calls hit 27%–99% of cases already rated untrustworthy; a forewarning cut execution in four of five models.
Δτ-Bench tests agents as tool ecosystems change
The strongest evaluated model or harness reached only 42.3% success on 1,588 tasks across six ecosystem states.
Trajectory2Gym builds decision-grounded supervision for tool-use agents
Across five Qwen3 scales it lifts the equal-weight Tau1, Tau2 and BFCL mean by 16.9 to 29.8 points.
NTBench shows higher success for near-target prompt injection
On language tasks, incorrect-target success rises 9.3 to 35.8 points versus a fixed-target baseline.
CRM-Agent-Bench tests tool-using agents on 555 CRM tasks
Sixteen systems score 64.32% to 83.60% strict success, with leaders diverging sharply in cost and domain reliability.
EpiSEM-Bench tests LLM agents when safety evidence is omitted
The benchmark has 300 matched unsafe/safe pairs across six domains and compares ten models.
AgentPolicyForge builds benchmarks and trains agent policy guards
On PolicyGym, a trained guard cut attack success 40-44% relative to prompting and rejected fewer benign requests.
Harness Scaling trains agentic RL across diverse harnesses
A factorized training subset improved transfer to unseen white-box setups and external black-box harnesses.
OpenVLHarness adds tools and memory for visual agent reasoning
On 22 benchmarks, average scores rise 9.7–12.6 points over matched no-tool baselines across eight backbones.
Retrieved PHREEQC output lowers GPT agent accuracy on fixed simulations
GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points via a section index, while two Claude models show no detectable drop.
MMedAgent-Plus coordinates 18 specialist models across medical tasks
End-to-end final-answer accuracy is 0.98, matching the dedicated specialist and above a 0.28 baseline.
OpenDS-World tests deep search agents on the live Web
The best of 12 frontier models scores 39.0%; humans score 21.1% across 1,000 attempts.
Heartbeat Thinking keeps voice-agent tools running while listening
On FDBench v3, tool coverage rises 16 percentage points and mean final-response latency falls 36.4 percent versus a windowed baseline.
Retrieved PHREEQC reports cost some GPT agents accuracy
GPT-5.1 and GPT-5.2 lost 11.3 and 7.9 points reading reports via a section index; Claude models did not.