AI papers — 2026-06-11

Problem: Standard MoE router rows are not aligned with the principal directions of their associated expert weight matrices, leading to suboptimal expert activation and load imbalance.

Method: Manifold Power Iteration (MPI) applies a "Power-then-Retract" update to each router row: one step of power iteration (R̂[i] = R[i]W_g^i(W_g^i)^T) followed by L2 normalization to a fixed norm constant. This aligns router rows with the principal singular direction of the corresponding expert. The method adds 0.2% training overhead and zero inference overhead.

Key results:

• Router-expert projection coefficient λ improves from ~0.25–0.37 (vanilla) to ~0.62–0.70 (MPI)
• Downstream accuracy +2.33 pp on 3B-parameter model, +1.84 pp on 11B-parameter model over 25 multiple-choice tasks (ARC, MMLU, GSM8K, MBPP, HellaSwag, WinoGrande, PIQA, DROP, SQuAD, CommonsenseQA, etc.)
• Pretraining loss reduced by 0.013 on MuonH-1B
• Load balancing MaxVio: 1.133→1.024 (batch), 0.964→0.711 (global) on 3B model
• Validated at 1B, 3B, and 11B scales (64–256 routed experts)
• Training data: FineWeb-Edu (350B tokens) + Olmo dataset (100B tokens)

Why it matters / Caveats: MPI offers a principled, near-zero-cost improvement to MoE router design grounded in spectral theory. Results are restricted to pretraining evaluations on multiple-choice tasks; downstream generation quality and applicability to already-trained models remain unstudied.

Problem: Autonomous research agents lack a persistent mechanism to accumulate, revise, and transfer knowledge across many experimental iterations within a long-horizon research session.

Method: Arbor separates autonomous research into a long-lived coordinator and short-lived executors. The coordinator maintains a Hypothesis Tree where each node stores a hypothesis, a distilled insight, and metadata (status, dev score, git branch). The coordinator loop follows Observe → Ideate → Select → Dispatch → Backpropagate → Decide, with each executor working in an isolated git worktree and returning a compact evidence report. Default budget is 20 coordinator cycles with a maximum tree depth of 2 and a 48-hour wall-clock limit.

Key results:

• Best held-out result on all 6 Autonomous Optimization (AO) tasks, achieving >2.5× the average relative held-out gain vs. Codex and Claude Code baselines
• BrowseComp: 67.67% accuracy (vs. 53.33% Claude Code, 50.00% Codex)
• Math-Reasoning Data Synthesis: 20.83 mean pass gap (vs. 8.33 Claude Code, 6.25 Codex)
• MLE-Bench Lite with GPT-5.5: 86.36% Any Medal, 77.27% Gold
• Ablation: removing the tree structure drops Any Medal from 81.82% to 63.64%; removing insight feedback drops to 54.54%
• Transfer: BrowseComp-optimized harness improved HLE from 25.50% to 31.50% and DeepSearchQA from 61.00% to 69.00%
• Token consumption: 20.12M–43.19M tokens per completed run, comparable to single-trajectory baselines

Why it matters / Caveats: Arbor demonstrates that structured hypothesis management with cross-run insight distillation substantially outperforms single-trajectory agents on diverse research tasks. All results rely on powerful proprietary LLM backbones (GPT-5.5, Gemini-3-Flash, Claude Opus 4.6), and cost at scale is not fully addressed.

Problem: The field lacks a unified framework for understanding how agentic environments for LLMs are designed, synthesized, evaluated, and used to train agents.

Method: This survey organizes agentic environment research around a four-phase lifecycle: modeling, synthesis, evaluation, and application. It identifies 8 environment attributes and 8 application domains. Synthesis paradigms are split into symbolic and neural approaches. Agent evolution in dynamic environments is characterized across four axes: memory-based experience, workflow orchestration, offline trajectory learning, and online exploration-driven development. Three environment evolution strategies are identified: neural-driven, difficulty-driven, and scaling-driven.

Key results:

• Not stated (survey paper; no experimental benchmarks reported)

Why it matters / Caveats: Provides a structured vocabulary and taxonomy for the rapidly growing area of agentic environment design. HTML version was unavailable (HTTP 404), so granular details about covered works and statistics could not be extracted beyond the abstract.

Problem: Existing SWE-bench evaluation frameworks impose technical requirements that prevent general-purpose agent harnesses (like OpenClaw) from being fairly assessed, and no benchmark separates the effect of model choice from harness design.

Method: Claw-SWE-Bench defines an Adapter Protocol that wraps any agent harness ("claw") into a SWE-bench-compatible interface via standardized lifecycle methods (create_agent, send_task, backup_session, delete_agent, get_docker_args), while a fixed orchestrator controls prompts, Docker runtime, 3600-second timeout, and patch extraction. The benchmark contains 350 real GitHub issue-resolution instances across 8 languages (Java, Go, Rust, JS/TS, C/C++, Ruby, PHP, Python) and 43 repositories. A Lite-80 subset (10 instances per language, 4 difficulty quartiles) reduces cost to ~22.9% of the full benchmark while maintaining 0.4 pp parity in Pass@1.

Key results:

• OpenClaw bare adapter (minimal direct-diff): 19.1% Pass@1; full adapter: 73.4% Pass@1 (same GLM 5.1 backbone) — a 54.3 pp gap from adapter design alone
• Model spread: GPT 5.5 78.0% Pass@1 ($1,399 cost) vs. Seed 2.0-mini 48.6% ($19.40) — 29.4 pp range
• Harness spread on Qwen 3.6-flash: 38.6%–66.0% across 5 claws — 27.4 pp range, exceeding neighboring model-tier differences
• Git future-commit leakage: Claude Opus 4.7 drops 84.7%→76.7% (−8.0 pp) when commits beyond base_commit are removed
• Lite-80 Spearman rank correlation with full-350: ρ=0.96; top-50 pipeline overlap: 42/50

Why it matters / Caveats: Demonstrates that harness engineering choices have as much impact on SWE-bench scores as model tier, which has direct implications for interpreting published leaderboard numbers. Costs (up to $1,399 per full run) limit accessibility for many researchers.

Problem: Scalar reward models for text-to-image generation fail to capture the inherently subjective and multi-dimensional nature of visual preference.

Method: Z-Reward is a two-model teacher-student framework. The 27B teacher (Qwen3.5-27B VLM) uses Group-wise Direct Score Optimization (GDSO) — combining policy-gradient rewards with direct pointwise and pairwise supervision — to produce score distributions over rubric-aligned bins via a Q-Align-style score decoder. The 9B student (Qwen3.5-9B) uses Reasoning-Internalized Score Distillation (RISD), compressing the teacher's reasoning-conditioned distribution into a compact model via KL divergence loss, eliminating explicit reasoning at inference time. Scores span a nine-level half-point scale (1.0–5.0) across four dimensions: Text–Image Alignment, Realism, Aesthetics, and Physical Plausibility.

Key results:

• 27B GDSO teacher: 89.6% human preference accuracy (HPA)
• 9B RISD student: 88.6% HPA (vs. 83.11% OPD baseline)
• PLCC improvement: +0.1319 (27B teacher), +0.3980 (9B student) over zero-shot VLM baseline
• Text-to-image optimization application: 41.3% net human-preference improvement over SFT baseline
• Evaluated on 400-prompt test set with internal human annotators

Why it matters / Caveats: Demonstrates that distributional reward modeling with reasoning distillation substantially outperforms scalar baselines for visual preference evaluation. Evaluation is on an internal dataset; generalization to other image generation systems requires further study.

Problem: Tabular encoders are evaluated inside task-specific end-to-end pipelines, making it impossible to fairly compare representations across different encoder paradigms.

Method: TRL-Bench decouples encoder evaluation into three suites sharing lightweight probe heads: TRL-CTbench (13 column/table-level tasks), TRL-Rbench (row-level prediction and record linkage), and TRL-DLTE (multi-stage data-lake table enrichment). Each encoder exports row-, column-, or table-level embeddings through a standardized wrapper, and probes are trained with table-disjoint train/dev/test splits. 20 models spanning 7 paradigms (generic text, table-text, table-structure, column-centric, prior-based, target-table SSL) are evaluated.

Key results:

• Benchmark assets: 50 OpenML tables with 123 targets, 16 record-linkage datasets, 47,772-table data lake from 1,379 parent tables (87 total datasets)
• BERT leads schema column-typing (F1: 0.926) and clean record linkage (F1: 0.418)
• TabICL leads row-level prediction (AUROC 0.816)
• Best DLTE pipeline: TUTA/GTE/GTE (0.229 UJ-H) vs. best monolithic BERT/BERT/BERT (0.139 UJ-H)
• Dev/test Spearman ρ=0.96; top-50 pipeline overlap: 42/50
• No single encoder wins universally; three identified gaps: specialization, transfer-scope, and composition

Why it matters / Caveats: Establishes that encoder capability is task-specific and that multi-encoder pipelines outperform any single model, changing the framing from "best tabular encoder" to "best encoder per capability type." The benchmark focuses on embeddings only, not full fine-tuning comparisons.

Problem: Existing code-agent training data focuses on bug fixing in existing repositories; there is no large-scale dataset for training agents to build complete repositories from scratch using only documentation.

Method: DeNovoSWE constructs 4,818 whole-repository generation instances via a sandboxed agentic pipeline requiring no human annotation. A Divide-and-Conquer workflow uses Overview and Capability agents to decompose the task, followed by a Draft-Critic-Repair cycle where iterative refinement drives quality. A Difficulty-Aware Trajectory Filtering strategy combines structural signals (log-normalized executable Python lines), LLM-based qualitative judgments, and empirical pass-rate estimates with weight optimization via grid-search Pearson correlation maximization. Models are fine-tuned on ~11K filtered trajectories generated by DeepSeek-V4-Pro High.

Key results:

• BeyondSWE-Doc2Repo: Qwen3-30B-A3B improves from 5.8% to 47.2% (+41.4 pp); Qwen3.5-35B-A3B from 43.8% to 50.0% (vs. GPT-5.4/CodeX at 61.7%)
• NL2RepoBench: DeNovoSWE-Agent-30A3B 23.0%, DeNovoSWE-Agent-35A3B 27.1%
• Dataset test coverage: mean 85.5%, median 89.6%, P90 at 99.9%
• Median repository: 79 unit tests across 12 test files, 9 source files
• DeNovoSWE is ~46× larger than NL2Repo (104 instances)

Why it matters / Caveats: Addresses a critical gap in long-horizon agentic code training data and demonstrates that fine-tuning on document-to-repository trajectories dramatically improves full-repo generation. Proprietary models (GPT-5.4) still lead, and the benchmark (BeyondSWE-Doc2Repo) contains only 50 test instances.

Problem: Egocentric video provides a limited, single-perspective view that causes MLLMs to make spatial reasoning errors that cannot be corrected without additional viewpoints.

Method: ReRe is a two-phase training-free inference framework. In the Reason phase, the MLLM forms spatial hypotheses from the source egocentric video. In the Re-reason phase, a Geometry-to-Video pipeline reconstructs 3D geometry with VGGT (Visual Geometry Grounded Transformer) and renders novel views along an Oblique Sweep Trajectory (diagonal camera path at 45° elevation), then passes these synthesized views back to the same MLLM for hypothesis revision. The pipeline requires no model fine-tuning and adds ~11 seconds per sample on an A100 GPU (~9s for VGGT, <1s for rendering).

Key results:

• VSI-Bench (5,000+ QA pairs from 288 egocentric videos): Qwen3-VL-4B 30.7→36.5 (+5.8%), Qwen3-VL-8B 30.5→35.8 (+5.2%), InternVL2.5-8B 35.5→36.7 (+1.2%)
• STI-Bench: Qwen3-VL-2B average +8.0, Spatial Relation subcategory +18.5
• Sample-level flip analysis: 71.6% positive flips vs. 28.4% negative (2.52:1 ratio)
• Oblique Sweep outperforms Mid-level Traverse (horizontal) and Bird's-eye Orbit (top-down) trajectories

Why it matters / Caveats: Shows that synthesized novel views can serve as cheap "second opinions" for spatial reasoning without any additional training. The pipeline depends on VGGT reconstruction quality and may degrade on scenes where 3D reconstruction is unreliable.

Problem: VLA models trained on static demonstrations struggle with dynamic, contact-rich manipulation scenarios because they lack explicit predictions of how actions will change the scene and how motion should proceed.

Method: World Pilot augments a VLA (Qwen3-VL backbone + DiT-based flow-matching action head) with two steering mechanisms fed by a frozen World-Action Model (Cosmos Policy, 5-step denoising). Latent Steering injects scene-evolution predictions into VLM hidden states via residual cross-attention. Action Steering compresses anticipated motion trajectories into a single prefix token prepended to the flow-matching action generator. A 0.3 dropout rate is applied to WAM conditioning during training to prevent over-reliance. The WAM remains frozen; only the VLA modules are trained.

Key results:

• LIBERO-Plus zero-shot OOD: 84.7% success rate (+4.2 pp over ABot-M0 baseline of 80.5%)
• Camera shift subset: 82.8% (+13.2 pp margin)
• Ablation: Latent Steering alone 83.7%, Action Steering alone 83.1%, combined 84.7%
• RoboCasa: 65.5% (competitive with baselines)
• Hardware: 8 RTX PRO 6000 GPUs for training
• World-model-only Cosmos-Predict baseline: 82.6% (+2.1 pp)

Why it matters / Caveats: Demonstrates that world-model predictions — even from video-trained models without robot-specific tuning — usefully steer VLA inference. The camera-shift improvement (+13.2 pp) is particularly strong. Results are limited to simulation (LIBERO-Plus, RoboCasa) and small real-robot task sets with 20 trials each.

Problem: Despite well-known computational disadvantages of Transformers, it is unclear which subquadratic architecture (xLSTM, Mamba-2, Gated DeltaNet) is superior and why.

Method: The paper compares xLSTM (variants: [7:1], [1:0], [3:1], [1:1]), Mamba-2, and Gated DeltaNet[-1,1] across three demanding application areas: code pretraining (400M-parameter models on Nemotron-CC-Code-v1, 20B–100B tokens), knowledge distillation from Qwen3-4B-Instruct (4B student models), and time-series foundation modeling (GIFT-Eval, 1M–80M parameters). Synthetic counting and state-tracking tasks (A^nB^n, Parity, Modular Arithmetic ℤ₅, S₃ word problems, Majority) at lengths up to 2048 tokens are used to isolate memory mechanisms.

Key results:

• Code pretraining HumanEval pass@64: xLSTM[7:1] leads by +1.43 pp at 20B tokens, +0.90 pp at 100B tokens, +1.81 pp on mixed corpus
• Code distillation (HumanEval/HumanEval+/MBPP/MBPP+ average): xLSTM[1:0] 0.768 vs. Gated DeltaNet 0.755 vs. Gated DeltaNet[-1,1] 0.756
• GIFT-Eval (10M params): xLSTM[3:1] 0.733 MASE / 0.508 CRPS vs. Mamba-2 0.767 / 0.525
• Synthetic Majority counting at length 2048: xLSTM[1:0] 0.763 vs. Gated DeltaNet 0.268 vs. Mamba-2 0.241
• Parity at length 2048: xLSTM[1:1] 1.000 (perfect) vs. Gated DeltaNet[-1,1] 0.472 vs. Mamba-2 0.352

Why it matters / Caveats: Provides a rigorous multi-domain empirical case for xLSTM's superiority, traced to superior gating-based memory correction mechanisms. All code experiments are at 400M/4B scale; whether advantages hold at frontier scales (70B+) is not tested.

Problem: Existing math benchmarks do not separately evaluate proof reasoning and constructive realization (producing a concrete witness object) at olympiad level, masking a fundamental capability gap in LLMs.

Method: ComBench contains 100 olympiad-level combinatorics problems (2000–2025) from 15 competitions (IMO, USAMO, and regional olympiads), split into 50 analysis-centric and 50 construction-centric records. Proof quality is graded with a 0/1/6/7-point IMO-style rubric by human annotators. Constructions are verified by deterministic Python verifiers checking discrete objects, and high proof scores are demoted if the construction verification fails (verifier-gated scoring). 14% overlap with IMO-Bench was identified and 86% of problems are novel.

Key results:

• Best overall: GPT-5.5 at 65.4% average, 75.3% Best@4
• Gemini-3.1-Pro: 60.3% overall; Kimi-K2.6: 53.5%; DeepSeek-V4-Pro: 45.2%
• Lower-tier models (Qwen3.6-Max, SU-01, GLM-5.1, etc.): 16.1%–24.9% overall
• GPT-5.5 construction pass rate: 75.0%; Kimi-K2.6 construction pass rate: 75.0%
• Hardest category: "Existence and Construction"; easiest: "Counting" and "Graph Theory"
• Dominant error type: "Missing Core Mechanism" (41.2% of failures)
• Capability separation confirmed: best proof model ≠ best construction model

Why it matters / Caveats: Reveals that proof competence and construction competence are distinct skills, and that even the best models top out at 75% construction success, indicating substantial headroom. Human grading of 100 proofs limits scalability for repeated leaderboard updates.

Problem: Grammar-Constrained Decoding (GCD), intended to improve code generation reliability, inadvertently removes a model's ability to produce natural-language refusals, enabling jailbreaks on safety-aligned LLMs.

Method: CodeSpear is an attack that applies a standard benign code grammar constraint during decoding, forcing the model's output space to exclude natural-language refusals and thus compelling generation of malicious code. CodeShield counters this with Direct Preference Optimization using "honeypot code" — semantically harmless but structurally valid code that the model learns to prefer over harmful code under GCD constraints. The preference hierarchy is: natural refusal > honeypot code (unconstrained) > harmful code (constrained). Training uses 2,000 malicious prompts from PKU-RLHF and honeypot code from OpenCodeInstruct (K=5 samples, 1 epoch, lr=1e-5).

Key results:

• CodeSpear average attack success rate (ASR) improvement: +26.90 pp over existing jailbreaks
• Average malware rate (MR) improvement: +17.98 pp
• Local deployment ASR: 81.82%; API-based ASR: 67.39%
• CodeShield reduces ASR to 5.57% and MR to 2.78%
• Utility degradation: HumanEval 70.93%→67.48% (−3.45 pp)
• 10 LLMs tested (Qwen2.5-Coder 7B/32B, Qwen2.5 7B/32B, LLaMA3-8B, GPT-5, GPT-5-mini, MiniMax-M2.5/M2.7, GPT-OSS-120B)
• Evaluated on RMCBench (182 requests), MalwareBench (320 requests), HumanEval (164 tasks), MBPP (974 tasks)

Why it matters / Caveats: Exposes a structural vulnerability in any GCD-based code generation system, not just specific models. CodeShield's small utility cost (−3.45 pp on HumanEval) makes it a practical defense. Human evaluation agreement was 87% (ASR) and 85% (MR), introducing some annotation uncertainty.

Problem: Existing open-source video models cannot process very long videos efficiently or operate as agents with tool use and memory, limiting long-horizon video understanding.

Method: InternVideo3 introduces Multimodal Contextual Reasoning (MCR), a closed-loop formulation where the model iterates over a shared evolving context containing observations, instructions, reasoning, tool actions, feedback, and memory. Efficiency at long contexts is addressed by Multimodal Multi-head Latent Attention (M²LA), which compresses KV-cache states using Low-Rank Latent Factorization, RoPE-Aware Positional Aggregation, and Modality-Aware Latent Adaptation, while reconstructing head-specific representations on the fly. Training is a four-stage pipeline: continued pretraining (CPT) on 16M samples (~13.5B tokens), short-to-long SFT (up to 2048 frames / 256K tokens), rule-based RL (GRPO on temporal grounding + MC-QA), and on-policy distillation from Qwen3-235B teacher.

Key results:

• Video-MME: 73.8 (best open-weight 8B); MLVU: 77.3 (best open); EgoSchema: 76.6 (best open)
• VRBench: 69.4 (best open); VSIBench: 68.1 (best open)
• QVHighlights: 59.9 (best open); Charades-STA: 50.4 (best open)
• M²LA decode throughput: 1.84× at 32K tokens, 4.12× at 128K tokens, 4.77–5.01× at 256K+ tokens
• M²LA enables 768K token contexts (base model OOMs at 512K); ~50% KV-cache reduction
• CPT ablation: −2.1 avg without CPT stage

Why it matters / Caveats: Delivers a best-in-class open-weight 8B video model while also serving as a functional video agent with hierarchical memory and tool routing. The agentic components (question routing, recursive verification) are evaluated qualitatively; quantitative agent-loop benchmarks are not centrally reported.

Problem: Multi-Token Prediction (MTP) speculative decoding acceptance rates collapse during RL training because the entropy increase of the policy under GRPO causes the draft head's token distribution to diverge from the current policy, invalidating standard greedy acceptance.

Method: Bebop identifies that MTP acceptance rate α is linearly negatively correlated with policy entropy H(p) when using target-only or CE/KL-trained rejection sampling (entropy-acceptance slope from −1.68 to −0.06 with TV training). The fix has two parts: replace greedy acceptance with probabilistic rejection sampling, and train draft heads with an end-to-end TV loss (e2e TV loss: 1 − (1/γ)∑_j ∏_i(1−d_TV(p_i,q_i))) that directly optimizes multi-step acceptance probability rather than per-token distribution matching. Pre-RL MTP training is sufficient; no online MTP updates during RL are needed.

Key results:

• Up to 95% MTP acceptance rates (agent tasks: 90.3%→97.0% with e2e TV)
• 1.8× end-to-end acceleration in async RL training
• Up to 2.4× per-step latency reduction on agentic tasks
• 25% extra inference throughput gains maintained throughout training
• SWE-Bench acceptance rate: 75.1%→83.1% with e2e TV
• Entropy-acceptance slope reduction: >95% (from −1.68 to −0.06)
• Evaluated on Qwen3.5-35A3B, Qwen3.6-35A3B, Qwen3.6-Plus, Qwen3.6-27B, Qwen3.7-Plus, Qwen3.7-Max
• Benchmarks: HMMT25, AIME25, LiveCodeBench, SWE-Verified, MT-Bench

Why it matters / Caveats: Solves a concrete bottleneck in RL training for LLMs by showing that standard MTP assumptions break under entropy increase and providing a theoretically grounded fix. Results are specific to Qwen-family models; generalization to other architectures and RL algorithms beyond GRPO is not tested.

Problem: Sparse autoencoders (SAEs) are computationally expensive to train and store, yet simpler classical methods like ICA have been dismissed as weak baselines for LLM interpretability without rigorous evaluation.

Method: ICALens is a GPU-parallel FastICA pipeline implemented in PyTorch with three stability improvements: row normalization (which increased accepted layers by 400% on GPT-2 Small), a p95-LIM convergence criterion (improved layer acceptance from 22 to 88 of 12 layers), and a minimum component count floor of 16. Activations are collected from 1 million post-block residual-stream tokens per layer using the Pile-10k corpus at 1,024-token context. Components are evaluated on SAEBench (sparse probing and Targeted Probe Perturbation) against public SAEs (Gemma Scope, Qwen Scope), Matryoshka SAEs, ITDA, PCA, and random projections.

Key results:

• ICA components show substantially higher excess kurtosis (non-Gaussianity) than SAE decoder directions and random projections across GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base
• Sparse probing: ICA competitive with public SAE dictionaries; outperforms PCA and ITDA consistently; Matryoshka SAEs fall below ICA
• Targeted Probe Perturbation: ICA strongest at small-to-medium budgets (5–20 features)
• Interpretability: 88.7% high-confidence human labels on 150 randomly sampled components (127/150 received non-unclear labels)
• Comparison types breakdown: word (54), phrase (33), sentence (20), form (11), other (32)
• FastICA total iterations reduced by 21.5% with stability recipes

Why it matters / Caveats: Rehabilitates ICA as a viable, training-free alternative to SAEs for LLM mechanistic interpretability, particularly for practitioners with limited compute. ICA produces fewer components than SAEs (not overcomplete), which may limit coverage of rare concepts; head-to-head coverage comparisons are not fully reported.

Problem: Reinforcement learning with verifiable rewards (RLVR) wastes rollout budget on prompts and intermediate reasoning steps that are either already solved or unsolvable, reducing reward contrast and training signal quality.

Method: TRACE (Tree Rollout Allocation for Contrastive Exploration) models multi-turn agentic rollouts as trees and allocates a fixed budget in two stages: (1) global root allocation prioritizes prompts where estimated success probability V̂_ψ(x_i, m) ≈ 0.5 (maximizing V_root = 1 − v_i^m − (1−v_i)^m), and (2) local prefix expansion allocates continuations to intermediate steps most likely to flip observed rewards. A shared predictor V̂_ψ is trained to estimate conditional success probability at both roots and prefixes. The framework is applied on top of GRPO with Qwen3-8B and Qwen3-14B backbones across math reasoning (DeepScaler), multi-hop QA (HotpotQA), and function calling (BFCL v4) tasks.

Key results:

• Multi-hop QA (HotpotQA/2WikiMultiHopQA/MuSiQue/Bamboogle): +2.8 pp average accuracy over GRPO at equal sampling cost
• Math reasoning: +1.1 pp (Qwen3-8B), +1.4 pp (Qwen3-14B) over GRPO
• Effective ratio (proportion of informative rollouts): 60.6% vs. 26.8% GRPO baseline (Qwen3-8B math)
• Baselines compared: ReAct, GRPO, PCL (prompt curriculum learning), TreePO

Why it matters / Caveats: Addresses sample efficiency in agentic RL training by targeting budget allocation to the most informative tree nodes, a complementary improvement to reward design. Gains are modest (1–3 pp) and experiments are limited to three task types with Qwen3 models; scaling and generalization to larger models or longer horizons are not assessed.