Self-HPO Lifts Qwen3.5-35B-A3B's Average Resolve Rate Across Three Coding Benchmarks From 57.8% to 64.8%

Today's Overview

Self-HPO Raises Qwen3.5-35B-A3B's Average Resolve Rate Across Three Coding Benchmarks From 57.8% to 64.8%. It alternates between updating the harness and policy. Switching to a stronger external harness proposer delivers a further boost.
Looped Models Need the Right Topology, Input Injection, and State Design. With matched training and inference compute, the full Huginn-style design uses 75% fewer resident parameters than a 112-layer feed-forward MoE. It performs better on DROP and GSM8K while tying on MATH500.
Synthetic Attribute-Binding Text Can Improve Visual Retrieval. In this study, fine-tuning an MLLM-based embedder exclusively on such text improved cross-modal and image-to-image retrieval while preserving general multimodal embedding performance. Open-world generalization still needs testing.

Featured

01 Better Models Still Need Better Harnesses

Coding agents depend on both the policy model and the harness that organizes prompts, tools, and environment interactions. When the policy improves, an old harness may still target yesterday’s failure modes. A fixed harness also constrains the training trajectories and what the policy can learn.

Self-HPO alternates between optimizing both components. The current policy analyzes its failures and proposes new harness candidates. Reinforcement learning then updates the policy under the selected harness. Training failures feed into the next harness search.

Each cycle changes the task and failure distribution, so joint optimization beats the sum of separate improvements. Across three coding benchmarks, Qwen3.5-35B-A3B’s average resolve rate rose from 57.8% to 64.8%. The gains carried into out-of-distribution tests, though broader generalization needs more evidence. Using the stronger external model Opus 4.8 as the harness proposer yielded a further performance boost, showing that Self-HPO still depends on proposer quality. Revisiting tool orchestration and interaction flows after a model upgrade may pay more than simply continuing to train the model.

Key takeaways:

Model upgrades change failure distributions, which can make an existing harness obsolete.
Alternating harness and policy optimization creates gains beyond improving each component separately.
Treat the execution framework as part of every agent upgrade, and track dependence on the external proposal model.

Source: Self-Evolving Coding Agents via Harness-Policy Optimization

02 Looped Models Depend on Three Design Choices

Reusing the same Transformer modules cuts resident parameters. Results still depend on three choices: loop placement, original-input injection, and recurrent-state organization.

The study isolates these variables while matching parameter scale, logical depth, and token budget. A topology with pre-loop, looped, and post-loop layers consistently improves context extraction and reasoning. Input injection mainly helps knowledge retrieval and context use. Randomly initialized recurrent states produce inconsistent gains. Separating the recurrent states can even hurt performance.

These choices also transfer to mixture-of-experts models. With matched training and inference compute, the full design uses 75% fewer resident parameters than a 112-layer feed-forward MoE. It performs better on DROP and GSM8K, ties on MATH500, and stays competitive elsewhere. Teams building small or sparse models should test topology and input injection before assuming more complex state mechanisms are an upgrade.

Key takeaways:

Looped-layer performance depends on topology, input injection, and state design, not parameter compression alone.
A three-part topology and input injection are the safest experiments to prioritize.
Cutting resident parameters by 75% preserves some reasoning strengths, but does not beat feed-forward models across the board.

Source: Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design

03 Synthetic Attribute-Binding Text Improves Visual Retrieval

Teaching attribute binding with only synthetic text can improve cross-modal and image-to-image retrieval. SABER covers four retrieval combinations across images and text. Its scenes contain as many as ten objects. CLIP-style dual encoders and newer MLLM embedders both struggle to bind attributes to the correct objects.

The researchers avoided image supervision. They fine-tuned an MLLM-based embedder exclusively on synthetic attribute-binding text with precisely controlled hard negatives. The resulting model reached the best current SABER performance while preserving general multimodal embedding quality.

Training also improved performance on negation and instance-retrieval benchmarks. The benefits of binding supervision therefore extend beyond SABER. Synthetic attribute-binding text can reduce the cost of collecting this supervision. Still, controlled benchmark gains do not guarantee open-world compositional understanding. The limits need more testing.

Key takeaways:

Attribute-to-object binding is a major failure point in multimodal retrieval.
Synthetic attribute-binding text can provide lower-cost supervision, and in this study its benefits transferred to visual retrieval.
Test the method in complex real-world scenes before treating controlled benchmark gains as general understanding.

Source: Stuck on Shuffle: On Attribute-Object Binding in Vision-Language Models

Self-HPO Lifts Qwen3.5-35B-A3B's Average Resolve Rate Across Three Coding Benchmarks From 57.8% to 64.8%

Also Worth Noting

Human Posters Become Verifiable HTML/CSS Templates Code IntelligenceRAP converts more than 25,000 human-designed paper posters into reusable slots and verifies outputs through rendering comparisons. Its held-out test set pairs 150 papers with manually designed posters. link.

Specialized Critics Cut 3D Scene Trial and Error Image GenSeparate design and critique experience trains a scene agent. A dedicated critic then supports both reinforcement learning and test-time correction. Scene-OTA-35K and Scene-OTC-15K provide design and quality-critique supervision. link.

Lean4 Verifies Rewards at Every Workflow Layer TrainingThe framework verifies tool calls, workflow stages, and final results with Lean4. Its stage-normalized GRPO trains LAR-9B and LAR-27B. The paper reports that LAR-27B reaches 77.3% on TerminalBench-2.1 and beats leading models of similar size. link.

One Interface Unifies Video Editing References Video GenA single reference-bank interface accepts variable numbers of text, image, and video inputs. It brings camera motion, subjects, style, and effects into one editing framework. R²-RoPE helps the model generalize to more references. link.

State Propagation Breaks Text World Models EvaluationWhen actions affect dependent objects, average task success across 20 models falls to 5.1%. GlobeBench contains 1,632 tasks across five domains and 19 simulated skills. No model solves 631 of them. link.

Harmless Self-Distillation Can Still Deepen Misalignment SafetyMisalignment can worsen through recursive self-distillation even when training prompts and model outputs are harmless. The effect appears in both the Qwen3 and Phi-4 families. Self-generated data loops need their own safety evaluations. link.

Deterministic Checks Protect Correct Formal Specifications ReasoningDRAGO diagnoses executable but semantically wrong specifications. It accepts revisions only when their error weight falls. This cuts damage to correct specifications from 11.8% to 1.3%. Across five benchmarks, answer accuracy beats zero-shot formalization by 12.9 percentage points. link.

Experience Cuts Coding Agent Token Costs Code IntelligenceCross-task experience supplies micro-skills, removes irrelevant tools, and stops failed loops without modifying the black-box agent. Average token costs fall by 18.4% to 30.9% while success rates stay similar across Claude Code, Codex, and GitHub Copilot CLI. link.

VISTA Tests Goal-Aware Egocentric Assistance MultimodalVision-language models can often describe first-person scenes but struggle to help users achieve their goals. VISTA tests that gap with six synchronized sensor inputs, about 1,000 egocentric recordings, and zero-shot evaluations of 14 representative models. link.

Answer Accuracy Hides Weak Audiovisual Grounding EvaluationModels with similar answer scores can differ sharply in whether annotated spatial, temporal, and audio evidence supports their explanations. ANCHOR contains 2,319 videos and 8,272 question-answer pairs. It separately measures correctness, hallucination, logical drift, and concision. link.

Monitor Transfer Treats Model Differences as Symmetries SafetyLEMON models attention-head permutations and depth differences as symmetries. Lightweight monitors can then transfer to new LLMs without fresh labels. Across 16 LLMs and six datasets, ROC-AUC beats the strongest baseline by as many as 14 points. link.