ISO: An RLVR-Native Optimization Stack

Reinforcement learning with verifiable rewards (RLVR) is rapidly advancing the reasoning capabilities of language models, yet the optimization layer that converts reward feedback into weight-space updates remains poorly understood. Building on our prior analysis (Zhu et al., 2025), we study this missing layer through the singular structure of model weights and identify spectral inheritance: RLVR can reuse the base model's weight spectra while acquiring new behavior through changes in the associated input and output singular frames. We operationalize spectral inheritance as Isospectral Optimization (ISO), an RLVR-native, fixed-spectrum optimization framework with complementary offline and online instantiations. Offline, ISO-Merger combines the frame changes of shared-base specialists into a single fixed-spectrum model, requiring no post-merge data, rollouts, gradient updates, or on-policy distillation (OPD). It recovers complementary specialist capabilities and achieves the strongest aggregate performance among the compared data-free merging methods. Online, ISO-Optimizer applies a chosen base optimizer, including AdamW and Muon, to the frame variables while keeping the base spectra fixed. Across reasoning and coding tasks ranging from 1.5B to 8B parameters, ISO-Optimizer improves accuracy in the reported runs and reaches matched scores with substantially fewer training steps. On Qwen3-8B-Base, AdamW reaches an aggregate accuracy of 0.495 after 270 training steps. ISO-AdamW reaches the same accuracy after only 100 training steps and improves further to 0.509 after 210 training steps. Together, ISO offers a concrete answer to RLVR's missing optimization layer: rather than inheriting pre-training optimization wholesale, design post-training around the structure of reward-driven adaptation: inherit the spectrum, optimize the frames.

Learning to Control LLM Agent Harnesses with Offline Reinforcement Learning

Large language model (LLM) agents are usually improved by changing prompts, models, or hand-written workflows, while the execution harness around the model is treated as fixed infrastructure. We argue that this harness is itself a learnable control layer. We formalize harness operation as a finite-horizon Harness MDP, where a lightweight controller selects structural execution actions while the LLM executor remains frozen. The controller is trained from offline rollouts using advantage-weighted regression with only terminal task-rubric rewards. We also separate final task quality from a post-hoc Harness Maturity Score, which measures whether the harness follows reliable execution patterns rather than only whether the final answer is correct. This separation gives a finite-buffer view of harness learning: final-quality gains require high-return support in the offline buffer, while process behavior can shift whenever it aligns with advantage-weighted actions. Across six controlled domains and two public-benchmark adapters, the learned controller consistently improves verification behavior and selectively improves final task quality, with the largest gains on adapted tau-bench retail, adapted AgentBench DB-Bench, and coding with a calibrated structural verifier. Ablations against behavior cloning and Forced CHECK show that the gains are not explained by imitation or by simply adding checks. These results identify harness control as a learnable layer for frozen LLM agents, while showing that offline support limits when better process control becomes better final answers.

Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents

Software-agent benchmarks usually report whether an agent solves a task, but the agent reaches that outcome through a harness that controls what it sees, which actions it can take, which failures are repaired, which states are verified, and which evidence is logged. We show that this harness can change the agent's multi-step beliefs even when the task, environment, and base LLM are fixed. We introduce a belief-rollout diagnostic that elicits structured K-step trajectories over progress, risk, recoverability, constraints, failure mode, uncertainty, future success, repair cost, and next action under alternative harnesses. We define a cross-harness belief divergence and decompose it into an arrival term for immediate interface shifts and a growth term for horizon-dependent belief changes. On controlled coding tasks and public-benchmark stress tests, blocked actions, compressed repairs, selective verification, and cost-aware evidence pruning often preserve terminal success while changing the beliefs that drive later decisions. We further introduce BIWM, a no-training protocol that canonicalizes observations, logs censored branches, expands repair traces, records verification masks, executes risky branches in shadow, and aligns belief trajectories across harness views. The results suggest that harness design is an experimental variable in agent evaluation, not an implementation detail. Our code is available at https://github.com/Hik289/Harness-induce-bias.git.

ManifoldFlow: SPD-Relaxed Stiefel Layers with Learnable Singular Spectrum

Orthogonal and Stiefel layers give neural weights exact spectral control, but they also impose a strong modeling constraint: all represented singular values are fixed at one. Many settings that benefit from an orthonormal basis still need direction-dependent attenuation or amplification. We introduce ManifoldFlow, a minimal relaxation of a fixed-spectrum Stiefel layer that keeps the basis on the Stiefel manifold while learning a bounded positive spectrum through W = Q S^{1/2}, with Q^T Q = I and S positive definite. Since W^T W = S, the eigenvalues of S are exactly the squared singular values of the realized weight, making eigenvalue clipping a direct singular-value control mechanism. Across paired sequence, tabular, and image experiments, the learnable SPD spectrum improves the fixed-spectrum Stiefel counterpart in the reported settings where the Stiefel prior is useful, with the largest gains in recurrent language-model projections. Boundary cases in convolutional classifier heads clarify the intended scope: ManifoldFlow is not a universal dense-layer replacement, but a spectrum-learnable Stiefel relaxation for settings where an orthonormal basis is a useful prior. When the basis should be orthonormal, its spectrum need not be frozen. Code available at https://github.com/Hik289/manifold_flow

Agent vs. Parametric World Models: Hybrid Planning for Reliable Language Agents

Language agents plan by generating not only actions but also implicit predictions of how the world will change. These imagined state updates make agents flexible, but they also create a distinct failure mode: hallucinated state claims can be written into context and propagated across subsequent decisions. In contrast, parametric world models provide measurable transition errors but are often weaker semantic planners. We study this tradeoff in graph-structured planning environments and introduce metrics for agent-world-model error, including hallucinated-state rate, propagation depth, and long-horizon error growth. We then propose Hybrid World-Model Planning (Hybrid-WM), which keeps the language model as the planner while using a small parametric transition model to predict action validity, state deltas, risk, and value. A consistency gate compares the agent's imagined delta with the parametric prediction and triggers targeted revision only under disagreement. Across four graph-structured planning benchmarks, Hybrid-WM improves success while reducing hallucinated state propagation. In live GPT-4o-mini evaluations, it reduces hallucinated-state rate from 0.176 to 0.035; in calibrated simulator ablations, it improves success from 0.668 to 0.838 with modest additional inference. These results suggest that lightweight parametric transition models can serve as effective grounding mechanisms for language-agent planning without replacing semantic reasoning.

Ask the World Before Acting: Environment Probing for Calibrated Agent World Models

Language agents acting over long horizons must maintain beliefs about tool states, object locations, graph edges, and subgoal dependencies. When these beliefs drift, failures can be fixed neither by longer reasoning traces nor by ordinary self-reflection, since the missing evidence lies in the environment. We formulate environment probing as a budgeted decision problem for structured agent world models: before acting, the agent may query the current value of one belief field, update its table, and pay one interaction step. We introduce EnvProbe, a simple scoring policy that combines task criticality, staleness, verbalized uncertainty, and dependency role. A type-stratified analysis separates the benefit of belief repair from the cost of displaced task actions and predicts different behavior for procedural and spatial beliefs. In three controlled environments with gold belief states, EnvProbe improves terminal world-state accuracy over periodic probing by 11.76 percentage points on procedural tool-dependency tasks, 3.79 points on spatial tasks, and 6.45 points overall. Ablations show that task-structural terms are the main source of the gains, while self-reported uncertainty is unreliable under confident wrong beliefs. The results suggest that agent calibration should be treated as an action-selection problem over environment evidence, not only as a model-internal reasoning problem.

Repair the Amplifier, Not the Symptom: Stable World-Model Correction for Agent Rollouts

As agent planning moves from short tool chains toward persistent workflows with thousands or tens of thousands of steps, failures will occur inside large planning graphs rather than in isolated predictions. Replanning the entire graph after every mistake is neither computationally realistic nor desirable: full-graph replay consumes large context budgets, exposes the LLM to many irrelevant symptoms, and can degrade long-context retrieval. This paper studies the missing component in such systems: a world-model corrector that repairs the failed planning graph in place. We compare two families of correctors. The first is the common engineering approach: scan nodes and edges, choose a suspicious local region, and ask an LLM to repair it. We implement strong engineering LLM correctors and find that they can help, especially when given very large contexts. The second family is our approach, WM-SAR (World-Model Subgraph Amplification Repair): instead of scanning for visible symptoms, it works backward from subgraph amplification, identifies the nodes and edges that keep re-amplifying error, and sends only that causal subgraph to the LLM. Across graph simulations and LLM repair experiments, WM-SAR substantially outperforms engineering correctors under realistic token budgets, achieves near-whole-graph stabilization with a compact region, and gives the LLM a cleaner repair target.

World-Model Collapse as a Phase Transition

Water looks unchanged as it warms, then at a critical point it boils. We ask whether long-horizon language agents show an analogous transition in their implicit world models. In some parameter settings, changing state load by a small amount, or adding a single step of horizon, leaves behavior nearly unchanged; near a critical boundary, the same small change causes a sudden world collapse. We study this effect in a deterministic task family with exact per-step gold state. A large grid search over state cardinality, dependency density, horizon, branching, observation mode, and mutation rate reveals a phase diagram: a solved plateau, a narrow transition band, and a collapse floor. Per-step traces show the mechanism: world-state fidelity fails before action validity, so the agent is not merely choosing a bad action; it is acting from a corrupted world. Stronger models translate the critical boundary but do not remove the qualitative transition. These results make world-model collapse a measurable bottleneck for long-horizon agents.

Ask the World Before Acting: Budgeted Environment Probing for World-Model Calibration

Long-horizon language agents do not only choose actions; they carry a private model of the world from one decision to the next. When that model drifts, a later failure can be decided before the failing action is ever taken. We study a direct repair mechanism: before committing to the next task action, an agent may ask the environment about one belief field and write the answer back into its world model. This makes environment interaction a scarce calibration resource, not merely a way to advance the task. We introduce \method, a budgeted probing operator for structured belief tables. The useful probes are not the same everywhere. Procedural beliefs, such as tool dependencies, can often be repaired by targeted checks, but those checks spend steps that the task may need. Spatial beliefs, such as object locations and graph edges, rely more on structural cues; the agent's own confidence can be a poor guide when the world changes off-screen. A type-stratified analysis formalizes this probe-action frontier, and controlled experiments show that mid-planning environment evidence reduces terminal world-model error when the probe policy follows the structure of the task.

AlgoBench: Benchmarking Algorithmic Adaptation in Code Generation

High pass rates on established programming benchmarks such as HumanEval and LiveCodeBench do not always show whether a model can reason about algorithms. Many fixed benchmarks eventually become part of the public training ecosystem through released problem statements, editorials, and generated solutions, allowing later models to improve partly by exposure rather than by stronger algorithmic ability. We introduce ALGOBENCH, a framework that automatically builds novel algorithmic problems from known competitive-programming problems through structured constraint-shifting transformations. Each accepted ALGOBENCH variant is traceable to a source problem, but must make the original reference algorithm fail. Beyond pass@$k$, we introduce complexity-aware metrics -- including OPTT, OPTS, TRAPRATE, GAPT, and CONSENS -- to test whether a solution is not only functionally correct but also asymptotically suitable for the generated problem. Experiments across multiple LLMs and prompting strategies show that performance drops sharply on ALGOBENCH variants, retrieval can increase reuse of the old algorithm, and many correct-looking solutions fail to meet the required complexity. Error analysis shows that failures are mainly algorithmic rather than implementation-level, suggesting that ALGOBENCH evaluates adaptation beyond functional correctness.