MemoryWalker addresses the conditioning problem that arises when training agents with compressed context during rollout. It introduces two exact, gradient‑equivalent corrections—LogitTree, a segmented K‑forward traversal, and a packed 4D attention mask—alongside SDCC, a self‑distillation method that reduces the train‑deployment gap by minimizing KL divergence at each eviction. Experiments on seven web‑search benchmarks show that SDCC significantly lowers logit drift and boosts rollout rewards compared to naive training.
arXiv:2604. 00004v2 Announce Type: replace-cross Abstract: The extension of context windows in Large Language Models is typically facilitated by scaling positional encodings followed by lightweight Continual Pre-Training (CPT).
LEGO-RL is a framework that connects native coding-agent harnesses with scalable policy‑gradient training without altering the harnesses’ internal flow. It achieves faithful optimization through in‑process LLM proxying, reliable execution via sandbox orchestration, and observable training with automated validation and a Live UI. Experiments show LEGO‑RL improves the Qwen3.5‑35B‑A3B model’s performance on three native harnesses while preserving high rollout‑training probability correlation.
arXiv:2607. 12463v1 Announce Type: new Abstract: Coding agents must integrate external tool returns into ongoing reasoning - a capability that standard left-to-right pretraining on code exposes only in its forward direction.
arXiv:2606. 05597v3 Announce Type: replace Abstract: Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary.
arXiv:2607. 15456v1 Announce Type: new Abstract: Looped, weight-tied Transformers reduce parameters by reusing a block, but decoding still stores a separate K/V cache for every recurrence step.
The paper addresses the memory bottleneck in reinforcement learning for large language models caused by the large Key-Value (KV) cache during rollout phases. It highlights that while KV cache compression can reduce memory usage, it introduces a significant off‑policy bias that standard statistical corrections cannot adequately mitigate. The authors argue that even tiny compression errors are amplified by RL’s instability, leading to inefficient learning.
arXiv:2608. 09447v1 Announce Type: cross Abstract: On-policy distillation (OPD) aligns a student with a teacher on trajectories sampled from the student itself, reducing the train-test state mismatch of offline distillation.
arXiv:2606. 08446v1 Announce Type: cross Abstract: Despite being powerful, reinforcement learning with verifiable rewards (RLVR) induces extremely long COT, making it computationally expensive.
The paper introduces Selective Supervision for Direct-OPD (S$^2$D-OPD), a refinement of Direct On-Policy Distillation that filters out states where the teacher’s policy change is minimal, as measured by the teacher‑reference Jensen‑Shannon divergence. By masking low‑divergence states and keeping only the top 10% of states per response, S$^2$D-OPD improves held‑out accuracy on AIME and HMMT benchmarks across multiple teacher‑student pairs without additional forward passes.
SCLATE is a new execution substrate that allows continual‑learning benchmarks and agents to share a single event scheduler via adapters, enabling tasks, session events, and memory consolidation to run on a compressed, real‑time timeline. It also functions as a rollout engine that records every model call’s tokens and log probabilities without modifying the agent’s harness or memory. Using SCLATE, the authors ported seven benchmarks, compared ten harness‑memory configurations across ten models, and demonstrated that post‑training Qwen3.5‑4B can effectively leverage both harness and memory, improving performance on multiple metrics.
arXiv:2607. 09375v1 Announce Type: new Abstract: We present Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters.
arXiv:2606. 05597v1 Announce Type: new Abstract: Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary.