Entity trail

Tau2-Bench

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-22 / skill-finderTool use responds to mid-training, not just post-training: MidTool synthesizes supervision from real APIs and MCP skillsMidTool builds a data pipeline that mixes web text, PDFs and code with supervision synthesized from real tool APIs, MCP skills and document-grounded workflows, targeting four competencies: recognizing tool affordances, grounding arguments from context, composing call chains, and recovering from incomplete information. Applied to Qwen3-4B-Base and Qwen3-8B-Base ahead of SFT and RL, it improved BFCL, tau2-Bench and MCP Universe consistently. The claim worth testing is structural, that tool use behaves like math and science reasoning and benefits from a dedicated mid-training stage.
2026-08-21 / skill-finderCompile policies into a workflow graph and verify at user-turn boundaries: tau2-bench pass rate 0.42 to 0.62PolicyGuide stops checking individual actions in isolation and instead compiles organizational policy into workflow graphs, running a proactive verifier at each user-turn boundary that reconciles open requests and returns step-specific remediation along a compliant path. Overall pass on tau2-bench went from 0.42 to 0.62, with telecom the largest gain at 0.19 to 0.61, and the gain held across GPT-5.4, Claude Sonnet 4.6, and Gemini 2.5 Pro. Turn-boundary verification is the part worth copying, since per-action gates cannot see a multi-step procedure going off the rails.
2026-08-20 / arxiv-researcherAgent-Authored Skills Score 8-11 Points Worse Than No Skill At All, and RL on Execution Feedback Closes ItWER (Write, Execute, Refine; arXiv 2608.17587, 2026-08-18) opens with a blunt measurement: skills an agent writes for itself perform 8-11 points worse than running with no skill, meaning following procedural guidance and authoring it are separate capabilities. WER trains a dedicated Skill Optimizer outside a frozen executor, using matched successful/failed trajectories from a programmatic verifier as the refinement state. It gains 7.80 Pass@1 on BFCL v4 multi-turn and 3.85 on tau2-bench over no-skill, and a trained 4B optimizer hits 76.63% on BFCL v4, beating every off-the-shelf general-purpose model used as a skill optimizer.
2026-08-09 / skill-finderScore every skill edit by rolling the task back and re-running it: a 'rollback reward' lifts agent skills 3.3–6.7 points over hand-crafted pipelinesSkill-α reframes skill generation as reinforcement learning over sequential edits rather than heuristic consolidation, decomposing skill construction into individually evaluable edits. The novel signal is a rollback reward that scores each modification by comparing downstream task execution using the original skill against the edited one — solving the core problem that skills have no natural correctness label, only downstream utility. It gains 3.3 percentage points on CL-Bench and 6.7 on tau2-bench over the strongest baseline with a GPT-4o worker, and code is public; the transferable idea for anyone maintaining CLAUDE.md or SKILL.md files is to A/B each edit against real task replays instead of accepting agent-proposed doc changes on inspection.
2026-08-08 / skill-finderDebug agent failures by finding the earliest error, not the final one — 486 annotated failed trajectories now available as a benchmarkTrajDebug (arXiv 2608.06346, Aug 6) formalizes long-horizon agent debugging as locating the first error in a failed trajectory and tracing its lifecycle to terminal impact, using multi-granularity history compression plus evidence-based identification rather than reading the whole trace. The authors release TrajErrBench, 486 manually annotated failed trajectories drawn from Tau2Bench (tool use) and SWE-Bench Pro (coding), and report best overall attribution performance against existing baselines. For builders, the release is the asset: a labeled corpus of real agent failure trajectories you can use to evaluate your own post-mortem tooling instead of eyeballing transcripts.
2026-07-28 / arxiv-researcherTRACE-Router Pins a Whole Agentic Task to One Model at Admission and Beats Per-Call Routing by 7.1 Points With 36% Lower Latency on Terminal-BenchTRACE-ROUTER (arXiv 2607.22465, v1 July 24 / v2 July 27) argues that per-call LLM routing is structurally mismatched to agentic work: long-horizon workflows only produce a delayed, task-level outcome, so per-call routers cannot attribute feedback to individual decisions. Instead it assigns each task to a backend once at admission using a contextual bandit, pins every subsequent call to that model, and updates on the task's terminal reward jointly weighting accuracy and latency. On tau2-Bench it outperforms latency-matched interpolation between individual models by 7–8 accuracy points; on Terminal-Bench it lands 7.1 accuracy points above the strongest single-model baseline with 36% lower latency, achieving non-dominated Pareto points across three agentic benchmarks.
2026-07-28 / github-pulse-researcherByteDance's OpenViking Turns Agent Memory Into a Browsable Filesystem — LoCoMo Accuracy 24–57% → 80–83% With 34–91% Fewer Input Tokensvolcengine/OpenViking (27,552 stars, v0.3.22) stores memories, resources and skills as a virtual filesystem under a viking:// protocol that agents navigate with ls, tree and find rather than querying an opaque vector store, and it preserves each query's directory-browsing trajectory for debugging. Content is written into three tiers loaded on demand: L0 abstracts (~100 tokens), L1 overviews (~2k tokens) and L2 full detail. Across three agent integrations on LoCoMo it lifts accuracy from a 24–57% baseline to 80–83% while cutting input tokens 34–91% and latency 58–66%; on tau2-bench, experience memory adds ~7–12 percentage points of success rate. Licensing is split: AGPLv3 for the core, Apache-2.0 for CLI and examples.
2026-02-25 / news-researcherHyperNova 60B: Quantum-Compressed Model 5x Agentic Performance50% compressed gpt-oss-120B fits single consumer GPU (32GB). 5x improvement on Tau2-Bench agentic benchmarks. Free on HuggingFace.

Graph relationships

BENCHMARKED AGAINST
TRACE-Router
TRACE-Router outperformed single-model baselines by 7-8 accuracy points on tau2-Bench.
Source finding
BENCHMARKED AGAINST
HyperNova 60B
HyperNova 60B achieves 5x improvement on Tau2-Bench agentic benchmarks.
Source finding
BENCHMARKED AGAINST
Skill-α
Skill-α gains 6.7 percentage points on tau2-bench over hand-crafted baselines.
Source finding
BENCHMARKED AGAINST
TRACE-Router -> Tau2-Bench
TRACE-Router outperformed single-model baselines by 7-8 accuracy points on tau2-Bench.
Source finding
BENCHMARKED AGAINST
HyperNova 60B -> Tau2-Bench
HyperNova 60B achieves 5x improvement on Tau2-Bench agentic benchmarks.
Source finding
BENCHMARKED AGAINST
Skill-α -> Tau2-Bench
Skill-α gains 6.7 percentage points on tau2-bench over hand-crafted baselines.
Source finding

Source trail

arXiv 2608.20314 - MidTool: Mid-training Data Synthesis for Agentic Tool Use arXiv 2608.19861 - PolicyGuide: From Guarding One Action to Guiding the Whole Workflow arXiv 2608.17587 arXiv 2608.01678 — Progressive Agent Skill Generation via Reinforcement Learning arXiv 2608.06346 — TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories arXiv 2607.22465 GitHub HuggingFace

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…