Learning What to Retain: Gated-Memory Routing for Efficient Collaboration in Multi-Agent LLM Systems

Large language model (LLM)-based multi-agent systems tackle complex reasoning by orchestrating how multiple agents are configured and how they collaborate. A central challenge is to adapt orchestration to the evolving collaboration state. Routing from the query alone cannot adapt to intermediate progress or errors, which hurts accuracy. Routing from the complete execution history supplies this missing context, but forces later decisions to process every prior step, including redundant or low-utility ones. This creates an execution-history overload that inflates cost. Effective orchestration instead requires a compact state that captures useful progress without accumulating redundant context. We propose Gated-Memory Routing, which conditions each decision on the query and a learned execution memory. A learned Memory Write Gate commits only non-redundant reasoning steps, and a learned Retrieval Gate supplies each agent a compact, relevant subset, so every decision conditions on a clean, informative state. At each step, the system selects the next role and backbone from this memory, while an Adaptive Halting Controller stops execution once the memory contains sufficient evidence for answering. Across five reasoning and code-generation benchmarks, our framework is both effective and efficient: it attains the best average accuracy, exceeding the strongest baseline by 2.44 points, while reducing HumanEval inference cost by 31.9% relative to that baseline. Code is available at https://github.com/rajibrhasan/gated-memory-routing

ACTD: Anchor-Based Cross-Tokenizer Distillation with Residual Regularization

Knowledge distillation effectively transfers reasoning capabilities from large language models to lightweight student models. To enable knowledge transfer across disparate model families, researchers increasingly explore cross-tokenizer distillation. However, cross-tokenizer distillation remains challenging due to vocabulary and sequence misalignment, while approximate vocabulary alignment can introduce additional noise into distillation. To address these challenges, we propose Anchor-Based Cross-Tokenizer Distillation with Residual Regularization (ACTD). ACTD bridges structural heterogeneity through vocabulary and sequence alignment, while mitigating alignment noise via a novel anchor loss with residual regularization. We further extend this framework to a multi-teacher setting. Evaluated across five reasoning benchmarks with three distinct teacher models, ACTD achieves state-of-the-art performance. Moreover, its multi-teacher extension outperforms the strongest single-teacher and multi-teacher baselines, further demonstrating the robustness of our method.

Dynamic Important Example Mining for Reinforcement Finetuning

Reinforcement fine-tuning (RFT) is increasingly used to strengthen the reasoning abilities of large models, yet its effectiveness is bound by how training data are selected and used. Most data-centric RFT methods rely on static or heuristic sample selection, implicitly assuming a sample's value is fixed over training. This overlooks the non-stationary dynamics of policy learning and can lead to suboptimal updates. We propose Dynamic Important Example Mining (DIEM), a principled and fully automated framework that makes data utilization adaptive throughout RFT. DIEM integrates two components into each optimization step: (i) a gradient-alignment importance estimator that efficiently approximates each sample's marginal contribution to policy improvement; and (ii) a constrained batch reweighting scheme that maximizes aggregate utility while preserving the update's gradient magnitude to stabilize optimization. Across several reasoning benchmarks, DIEM consistently outperforms strong static and dynamic baselines. The code will be released via https://github.com/hrtan/DIEM.

Prefix-Denoising Consistency: Test-Time Verification for Diffusion Language Models

Diffusion Language Models (DLMs) have recently become increasingly competitive with autoregressive (AR) models, and even outperform them on certain tasks. Unlike AR models, DLMs produce output through iterative denoising without a left-to-right order. To further improve the performance of DLMs, we introduce PDC (\emph{Prefix-Denoising Consistency}), a test-time self-verification method for DLMs. PDC exploits a distinctive test-time signal in DLMs under prefix conditioned regeneration, correct trajectories are more stable and reproducible than incorrect ones. Concretely, given an initially generated sample, PDC splits the sentence at an intermediate position and regenerates the remaining tokens conditioned on the fixed prefix. Across mathematical reasoning and commonsense reasoning benchmarks, PDC consistently improves upon the initial sample, outperforms independent generations under a computational constrained comparison, and is robust to different unmasking strategies and parameter settings. These results highlight prefix-conditioned regeneration as an effective DLM-specific primitive for test-time verification.

Understanding Curriculum Learning in Large Language Models via Cross-Difficulty Optimization Dynamics

Curriculum learning has been widely adopted in the post-training of large language models by organizing training data from easy to hard. However, its effectiveness varies substantially across reasoning tasks, suggesting that no single curriculum is universally optimal and raising a fundamental question: what determines when curriculum learning works? In this paper, we answer this question by analyzing the optimization dynamics induced by different curriculum schedules. We show that the transfer relationship between different difficulty levels characterizes the optimization dynamics induced by curriculum learning, which in turn explains the effectiveness of different curriculum schedules, and formalize this relationship as Relative Transfer, a principled measure of cross-difficulty knowledge transfer. Based on this measurement, we derive Transfer-aware Dynamic Curriculum Sampling (TDCS), which dynamically adjusts the sampling distribution according to the estimated transfer relationship throughout training. Extensive experiments on multiple reasoning benchmarks demonstrate that TDCS consistently outperforms representative scheduling strategies across different tasks, model scales, and training paradigms. More importantly, our work provides a unified optimization-based explanation of curriculum learning through cross-difficulty transfer.

GRIP: Grounded Reasoning via Information-Restricted Premises

High-capacity encoders in retrieval-augmented generation (RAG) can let the query dominate the latent state, leaving retrieved evidence functionally irrelevant. We call this failure mode query dominance. To address it, we introduce \textbf{GRIP} (Grounded Reasoning via Information-Restricted Premises), which imposes capacity asymmetry: the decoder keeps full-dimensional access to the query, while retrieved evidence passes through a severe stochastic bottleneck. This forces the evidence channel to encode only the residual information unavailable from the query. Across five reasoning benchmarks, GRIP outperforms strong iterative baselines, cuts a query--latent mutual-information diagnostic by roughly 30$\times$ (14.8 $\to$ 0.47 bits), and reduces hallucination by 73\%. Residual-alignment analysis further shows that the bottleneck output occupies subspaces less aligned with the query than baseline representations.

Who Thinks Best Depends on How Long You Let Them: Budget-Dependent Rankings in LLM Evaluation

Standard evaluation of large language models assumes stable model rankings across inference conditions. We challenge this assumption by varying the token generation budget, i.e., the maximum tokens a model may produce, across seven levels (64--4,096), evaluating four models on three reasoning benchmarks (56,476 inferences). We report four findings: (i) 3--19% of items exhibit non-monotone behavior (accuracy decreasing with more budget), even after controlling for truncation, and this phenomenon is model-specific (cross-model overlap: 6--14%). (ii) Model rankings reverse across budgets on all benchmarks ($p {<} 0.01$, McNemar). (iii) Oracle analysis reveals model complementarity up to $+27.8$pp, most pronounced at constrained budgets. (iv) A budget-aware router captures 14.1% of the oracle gap cross-domain; budget features help within-domain ($+1.6$ to $+5.7$pp) but are domain-specific and hurt transfer ($-1.2$pp). These results argue for budget-conditioned evaluation protocols.

Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility

Large language models can solve substantially harder reasoning problems with more inference-time compute. The term "test-time scaling," however, now covers diverse inference algorithms that extend deliberation along a single trajectory, sample completed candidates and aggregate them through voting or verification, or search over unfinished partial states. These algorithms differ in their statistical structure, compute accounting, and failure modes. Treating these procedures as interchangeable under a single scalar "budget," or reporting accuracy without the inference protocol that produced it, makes results difficult to compare across studies. We develop a systematic account of test-time scaling along three axes. First, we formalize test-time scaling as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish three structural regimes: single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the evaluated object as the entire inference system and develop evaluation principles that separate end-to-end system performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and prescribe protocol-matched reporting of compute and uncertainty. Third, we specify reproducibility requirements for inference protocols, distinguishing exact replay from distributional reproducibility and identifying the artifacts needed to support each. We also organize the open-weight reasoning ecosystem by model-side and interface mechanisms, apply these principles to broad-knowledge, symbolic-reasoning, and competition-mathematics benchmarks, and assemble over 2 billion full reasoning traces for release with progressively richer verifier and token-level signals.

SKIMIX: Multi-Agent Harness-Time Scaling with Skill Mixture for Dynamic Harness Engineering

AI agents increasingly rely on large skill libraries, but selecting, combining, and maintaining skills remains difficult. We propose SKIMIX, a multi-agent framework in which agents with different skill portfolios collaborate through iterative refinement. SKIMIX combines embedding-based skill retrieval, submodular anti-dilution routing, and adaptive skill evolution. Across six reasoning benchmarks, multi-agent collaboration substantially improves open-ended mathematical reasoning but offers limited or negative gains on multiple-choice tasks. Agent-count scaling is non-monotonic, and most improvements arise during the first refinement round. These results show that task characteristics determine whether skill-level ensembles help and provide practical guidance for scalable agent design.

Frustratingly Simple Black-Box Adaptation of Language Models via Logit Bias

Many organizations aim to adapt language models for internal use, both to improve performance on domain-specific tasks and to address privacy concerns around sensitive data. However, such adaptation remains non-trivial: it often requires operationally challenging fine-tuning of open-source models or ad hoc prompt optimization. We study a minimal alternative based on a simple API-level control: allowing users to bias the model's logits with a user-defined vector. We develop a black-box method for learning a single context-independent logit-bias vector, added at every decoding step, without modifying model weights or requiring gradients. Starting from a KL-regularized reinforcement learning (RL) objective, we characterize when such a fixed logit-bias vector can approximate the optimal prefix-dependent correction and derive a closed-form inverse-propensity estimator from rollouts, rewards, and token probabilities. Empirically, this simple decoding-time intervention improves over base models on mathematical and reasoning benchmarks while using far fewer trainable parameters than conventional fine-tuning. Our results suggest that learned logit bias is a lightweight mechanism for adapting language models under minimal access requirements.

QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling

Scaling inference compute, by generating many parallel attempts per problem, is a costly but reliable lever for improving language model capabilities. By default these attempts are generated independently, wasting inference compute on redundant solutions. This waste seems unavoidable. After all, independence is what makes parallel sampling trivial to scale. However, this tradeoff is not fundamental: there is a rich design space of samplers that generate correlated but exact samples entirely in parallel. We explore this design space as an avenue for improving sample efficiency in scaling inference compute and reinforcement learning (RL). Concretely, we introduce QuasiMoTTo, which uses correlated samples as a drop-in replacement for i.i.d. samples. To generate these samples, QuasiMoTTo uses a reparameterization of autoregressive sampling as inverse-CDF sampling and draws the underlying uniforms with quasi-Monte Carlo (QMC); because QMC spreads the uniforms out more evenly than i.i.d., the resulting samples cover the output space with far less redundancy. Even though the batch is correlated, each sample is marginally distributed according to the language model, so we can use the batch for policy-gradient training. Our empirical analysis focuses on understanding how efficiently QuasiMoTTo can turn compute into performance. To evaluate correlated samplers, whose dependence breaks standard pass@k estimators, we first develop an unbiased bootstrap estimator. Across four reasoning benchmarks, QuasiMoTTo matches i.i.d. pass@k accuracy with 25-47% fewer samples. Strikingly, QuasiMoTTo often saturates an upper bound on pass@k that holds for any marginal-preserving sampler. We also apply QuasiMoTTo to policy-gradient RL (GRPO) where it matches i.i.d. performance with 50% fewer training steps. These gains come from higher coverage, which yields a stronger learning signal per batch.

Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding

Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring Guess-Refine-Perturb dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer through entropy-guided conservative backward search. We further provide a theoretical formulation of layer selection as an optimal stopping problem, showing that under bounded projection noise and dominant late-stage alignment perturbation, our search rule filters perturbation while bounding the loss relative to the oracle refinement layer. Experiments across dense and Mixture-of-Experts LLMs demonstrate consistent gains on challenging reasoning benchmarks, including GPQA-Diamond, Omni-MATH, and HLE, with zero memory overhead and less than 2% latency increase. These results suggest dynamically bypassing final-layer perturbations can unlock stronger reasoning behavior from aligned LLMs.

Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers

Looped architectures provide an inductive bias toward learning step-by-step procedures for tasks that require compositional reasoning. The number of effective layers reached by looping determines the quality of the solution these models find. Like deep architectures, looped architectures are prone to a signal propagation problem induced by depth as the halting decision is postponed. In this paper, we address this signal propagation issue using pre-norm layers and residual scaling. Building on these architectural modifications, we propose FPRM, a Transformer-based Fixed-Point Reasoning Model that uses fixed-point convergence as an end-to-end halting mechanism in a looped architecture. We show that fixed-point halting allows FPRM to adapt its compute to task difficulty. FPRM is effective on common reasoning benchmarks, namely Sudoku, Maze, state-tracking, and ARC-AGI.

Learning from the Self-future: On-policy Self-distillation for dLLMs

On-policy self-distillation (OPSD) has proven effective for post-training large language models (LLMs), yet its application to diffusion LLMs (dLLMs) remains unexplored. Existing OPSD methods are inherently autoregressive-centric. They inject privileged information via left-to-right prefix conditioning with token-level divergence supervision, a design that fundamentally conflicts with the arbitraryorder generation of dLLMs. We introduce d-OPSD, the first OPSD framework tailored for dLLMs. Our approach makes two core contributions. First, we reframe self-teacher construction by using self-generated answers as suffix conditioning, enabling the student model to learn from "self future-experience" rather than privileged prefixes. Second, we shift supervision from token-level to step-level, aligning training with the iterative denoising process of dLLMs. Experiments across four reasoning benchmarks show that d-OPSD consistently outperforms RLVR and SFT baselines with superior sample efficiency, requiring only around 10% of the optimization steps by RLVR and opening a promising pathway for dLLM posttraining. The code is available at https://github.com/xingzhejun/d-OPSD.

Understanding Diversity Collapse in RLVR via the Lens of Overtraining

Reinforcement learning with verifiable rewards (RLVR) has become a key approach for enhancing the reasoning abilities of large language models. However, RLVR often suffers from \emph{diversity collapse}: Pass@$1$ improves while high-$k$ Pass@$k$ degrades, which is viewed as a narrowing of the model's reasoning boundary. We formalize this diversity collapse through the lens of \emph{overtraining}: once a problem's contribution to the reference metric has effectively saturated, further updates no longer expand what the model can solve but still concentrate probability mass on the trajectories favored by on-policy sampling. Under a standard setup with few rollouts per problem, even a single observed success places a problem in a nearly saturated regime for high-$k$ Pass@$k$, so most updates in standard RLVR are overtraining from the boundary perspective. This perspective also suggests a reading of whether RLVR can expand the model's reasoning abilities beyond the base model: since RLVR is structurally biased against high-$k$ Pass@$k$, its aggregate decline does not by itself mean that no new reasoning gains occurred. Interventionally, restricting updates to problems with zero observed success lifts Pass@$256$ above the base model on difficult benchmarks; observationally, a non-trivial fraction of initially unsolvable problems become solvable during standard RLVR training. Building on these findings, we propose \emph{Bayesian Boundary Gating} (BBG), which redirects optimization away from overtraining by estimating each problem's marginal contribution to the reasoning boundary. Across multiple reasoning benchmarks, BBG improves average Pass@$k$ across a wide range of $k$.

Forward-Free Diffusion Language Models with BPTT-Free Looped Refinement

Diffusion language models generate text through iterative denoising, offering a powerful alternative to autoregressive generation. However, discrete language spaces lack a natural neighborhood structure for defining effective perturbations, motivating artificial corruption schemes in the forward process. Such prescribed forward processes often produce states that are mathematically convenient but misaligned with the drafts and errors encountered during generation, resulting in degraded sample quality. To address this limitation, we propose FReDA, a forward-free diffusion language model that eliminates the need for a hand-designed forward process. We formulate diffusion language modeling as recursive distribution refinement, in which model-generated drafts serve as implicit intermediate states and the learned refinement model progressively moves the draft distribution toward the target distribution. Training detaches the preceding refinement passes and backpropagates only through the final pass, avoiding backpropagation through time (BPTT) across refinement iterations while retaining standard backpropagation within the final Transformer pass. Concretely, FReDA refines drafts by proposing candidate sequences and either directly performing self-refinement or selecting among parallel candidates via Best-of-N refinement. With this design, FReDA is neighborhood-agnostic, sampler-adaptive, and compatible with flexible refinement parameterizations. Extensive evaluations in the sub-8B regime show that FReDA-4B outperforms larger diffusion base models on reasoning and coding benchmarks, achieving absolute gains of up to 15%, while reaching a 1.5-1.8x average speedup over diffusion baselines and scaling effectively with additional refinement computation.

DLLG: Dynamic Logit-Level Gating of LLM Experts

Leveraging multiple specialized LLMs can combine complementary strengths, but existing approaches trade adaptability for stability: routing commits prematurely, heuristic ensembling depends on fragile proxies, and parameter merging introduces interference. We propose DLLG (Dynamic Logit-Level Gating), a dynamic logit-level ensembling framework that learns token-level expert fusion from sparse response-level supervision. A lightweight gating module predicts step-wise fusion weights, linking trajectory-level correctness to generation without token-level labels or expert retraining. Across diverse reasoning and code benchmarks, DLLG consistently outperforms strong routing, heuristic ensembling, and parameter-merging baselines across model scales, highlighting learned logit-level fusion as a robust and scalable paradigm for integrating specialized experts.

Tool-Aware Optimization with Entropy Guidance for Efficient Agentic Reinforcement Learning

Agentic reinforcement learning (RL) equips large language models (LLMs) with tool-use capabilities that substantially improve reasoning on complex tasks. However, integrating external tools often destabilizes training: over-reliance on tools can induce input distribution shift, while overly conservative tool use limits effective exploration. To address this issue, we propose a unified framework TAO-RL that couples tool-aware trajectory filtering with entropy-guided exploration for efficient policy optimization. Specifically, at the data level, TAO-RL filters rollout trajectories along two criteria: discarding those where all tool invocations fail to execute, and removing those where all rollouts are either correct or incorrect, as both cases yield degenerate advantage estimates that contribute no discriminative learning signal. This joint filtering retains data that are both tool-capable and informative, establishing a high-quality training distribution. At the algorithmic level, we introduce a tool-aware entropy-guided bonus that reshapes the advantage function at post-tool-call tokens, encouraging the policy to explore more diverse reasoning paths at critical decision points. These two components are mutually reinforcing: trajectory filtering establishes a clean and informative training foundation, while entropy-guided exploration drives stronger reasoning behaviors at critical tool-interaction junctures. Extensive experiments on 7 challenging reasoning benchmarks across 3 model scales demonstrate the superiority of TAO-RL over existing methods.

Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) plays a pivotal role in improving the reasoning ability of large language models. However, widely used PPO surrogate objectives are fundamentally local, as they rely on a local approximation of the exact policy gradient objective. While this approximation improves stability by reducing the variance induced by importance sampling, it also introduces structural bias into the surrogate objective, which must be controlled through trust region mechanisms. In this work, we introduce the $N$-step forward trace, which augments the PPO surrogate objective using the cumulative likelihood ratio of the next $N-1$ tokens. Building on this idea, we propose $N$-Step Forward-Trace Policy Optimization (NFPO), a practical RLVR algorithm that integrates the $N$-step forward trace into the masked policy gradient framework. NFPO provides a continuous bridge between the PPO surrogate objective and the exact policy gradient objective, offering a principled mechanism for controlling the bias-variance trade-off. Our theoretical analysis shows that, with an appropriate choice of $N$, the proposed objective yields a tighter policy-improvement bound than the standard PPO surrogate. Experiments on comprehensive reasoning benchmarks demonstrate that NFPO consistently improves performance, supporting our theoretical findings.