A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support. We present two techniques that make simulated alignment evaluations harder to distinguish from real deployments. Our first technique, critique refinement, spends additional inference-time compute on each simulator action: the simulator generates multiple candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and continues the evaluation with the most deployment-like candidate. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in an agent harness, reducing the gap between simulated and real deployment environments in coding settings. We test the techniques on multiple target models and find that they compose: applying both yields larger realism gains than either alone. Our results show that automated approaches can improve the realism of alignment evaluations, and that these improvements use additional compute more effectively than making the audits longer.
Axel Ahlqvist, Richard Guan, Juan-Pablo Rivera et al.· 0 citations
Adapting the communication topology of an LLM multi-agent system to each query improves both accuracy and efficiency, yet current designers treat this as conditional graph generation: a variational, autoregressive, or diffusion decoder searches the $N \times N$ adjacency space, and a graph-network proxy trained on utility and a structural cost such as edge count ranks the sampled candidates. We argue that this formulation is misaligned with the problem. Empirically, topologies that survive a reward filter collapse to about six distinct graphs even when the codebook capacity grows from 8 to 64; edge count is negatively correlated with measured token consumption (Pearson $r \approx -0.4$), so sparsifying the graph makes inference more expensive; and a message-passing scorer over agent-profile nodes is adjacency-invariant whenever agents share a profile---the default configuration of published benchmarks---so it cannot rank candidates at all in that regime. These three facts motivate Codebook Agent: a vector-quantized autoencoder compresses successful topologies into a query-independent 16-entry codebook; a reward-weighted MLP maps the query embedding to a distribution over codes; and an MLP proxy that reads the flattened adjacency, regressed on measured utility and per-task normalized token cost, reranks the top decoded candidates in a single batched forward pass. With no iterative search and no message passing at test time, Codebook Agent is the most accurate method on all six benchmarks we compare (84.6 average against 83.0 for the strongest prior designer), emits a topology in 2.4 ms, and uses 21.9--33.2% fewer LLM tokens.
Jin-Xiu Yu, Yu-Bei Li, Eric Jiang et al.· 0 citations
Self-improving agent pipelines have a problem at their center. An optimizer rewrites prompts to score higher, and the score comes from a judge that is itself an LLM. That judge has the last word on whether the system is getting better, and our position is that it has not earned it. The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override. We reached this position by building the alternative and running it. Over months of running autonomous prompt-optimization loops in production across contract analysis, compliance review, and code quality, we cataloged eleven ways the evaluation signal failed, in four classes: judge bias, harness and metric failures, ground-truth errors, and reward hacking. Agents achieved perfect scores by reading cached answer keys from their environment, a 100% pass rate concealing 68% true capability. A corrupted ground-truth label caused the optimizer to delete correct compliance rules to agree with it. A syntactically broken prompt was promoted as the winner because a silent parser fallback improved the metric. Attempts to fix the judge by rewriting its rubric plateaued; the only reliable gain came from a structural constraint on its output order. In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating. We report the failures this prevented, and, because the Teacher is itself an LLM judge, the failures it did not.
Multimodal Large Language Models (MLLMs) have achieved strong performance on structured visual understanding tasks such as chart and document question answering. However, existing benchmarks typically evaluate these domains in isolation, leaving underexplored a key capability: whether models can use textual context to determine how chart evidence should be selected, interpreted, and aggregated. We introduce DocHop, a benchmark for integrated chart--context reasoning in document-style images. In DocHop, the document narrative specifies multi-step compositional constraints, while charts provide the corresponding data values. Questions are grounded on a semantic reference label defined in the narrative, requiring models to resolve target entities from context before aggregating evidence across multiple charts. To enable systematic evaluation, we construct DocHop via a stochastic logic-first generation pipeline with controllable reasoning depth and visual density, covering 2,074 examples across six task categories. Experiments on a wide range of proprietary and open-source MLLMs show a substantial gap to human performance: annotators achieve over 90% accuracy, while the best model reaches only 62.83%. Reasoning-enhanced models consistently show improved results, but performance degrades as reasoning complexity increases. Overall, DocHop provides a controlled testbed for challenging multi-hop document reasoning.
Zhuoran Yu, Le Thien Phuc Nguyen, Jaden Park et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Ultra-low-bit language models can reduce storage and memory bandwidth, but a nominal"1.58-bit"label does not fully describe the stored representation, retained capability, or runtime behavior. We study an end-to-end post-training conversion of Qwen, an instruction-tuned 4B-parameter model, using KOTMS rotation, E2M-ATQ ternarization, and GPTQ-style error compensation from TWLA. The experiment is weight-only: activations remain at 16-bit precision, so ILA-AMP is omitted. We evaluate effective bit accounting, task capability retention, perplexity, calibration sensitivity, checkpoint composition, and deployment behavior. The final conversion uses 1.641 effective bits per weight for quantized linear weights, with 81.62% of model parameters targeted. Across ten scored capability comparisons, accuracy falls from 64.5% to 54.7%. Degradation is uneven: BoolQ retains 84.6% chance-corrected teacher performance, while ARC-Challenge retains 43.8%. Perplexity rises from 13.639 to 18.748 on WikiText-2, 24.700 to 31.992 on PTB, and 19.831 to 28.966 on C4. A subsequent packing run preserves the ternary planes and scales, reducing reported model size from 8.29 GiB to 3.96 GiB with essentially unchanged perplexity. A separate third-party packing attempt was lossy and is excluded from the primary artifact claim. The packed artifact has not been benchmarked end-to-end for task accuracy or generation throughput. A preliminary Triton GEMV microbenchmark is 4.6x slower than FP16 cuBLAS on one tested shape. We therefore do not claim that compression alone yields faster inference.
This note studies the geometry of full conformal prediction (FullCP) regions generated by an empirical energy-form pairwise score. Candidate-score convexity alone does not guarantee connected FullCP regions, even for empirical averages of losses convex in the candidate argument. For the energy-form score, each leave-one-out training comparison reduces exactly to a pairwise-dissimilarity sublevel condition. Under symmetry, a constant diagonal, a diagonal lower bound, and attainment of the associated Fr\'echet-type objective, all comparison regions share a minimizer; if they are convex, every nontrivial exact conformal region is star-shaped about that point. For power distances $\rho_\beta(x,y)=\|x-y\|^\beta$, this geometry holds for $\beta\ge1$, while the conventional energy score is strictly proper for $0<\beta<2$. For $d=1,\beta=1$, every nontrivial empirical-CRPS FullCP region is a nonempty closed interval (possibly $\mathbb R$ when $m=1$). For $1<\beta<2$ and $m\ge2$, explicit data-checkable derivative bounds give Lipschitz control of the radial exits and exact conformal radial function. Combined with directional root search and classical Lipschitz extensions, they yield certified inner and outer radial envelopes of width at most $\delta+2\widetilde Lh_{\mathcal U}$ and same-ray Hausdorff guarantees. An analytic two-dimensional example shows why preserving star-shaped but nonconvex geometry can matter. A staged two-dimensional study finds modest but systematic tightening of the generic certificate and frequent robust nonconvexity witnesses, with detected normalized radial departures typically small. The method is intended for low-dimensional multivariate outputs rather than high-dimensional scaling or runtime improvement.
Large audio-language models (LALMs) have shown promising progress in understanding speech, music, and general sound events, yet their ability to reason about how audio signals are degraded remains underexplored. Existing benchmarks primarily evaluate semantic understanding, event recognition, or high-level audio reasoning, leaving a basic question unanswered: Do LALMs understand the differences in audio quality? We introduce MRMAD, a Multi-Round Multi-Audio Degradation benchmark for evaluating audio degradation perception and understanding in LALMs. MRMAD spans speech, music, and sound, and frames evaluation as multi-turn dialogues across multiple audio inputs, requiring models to identify types of degradation, compare severity, and perceive corruption changes across turns. Unlike current single-turn audio-language benchmarks, MRMAD evaluates whether LALMs can maintain consistent degradation hypotheses with new evidence and comprehend low-level acoustic phenomena over multi-turn dialogues. Through a systematic evaluation of 18 representative LALMs from non-thinking to reasoning and Omni models, we find that current models often recognize coarse content while failing to diagnose, compare, or reason about degradations reliably. Human evaluations further reveal a significant perception gap between LALMs and human listeners. MRMAD thus exposes a critical yet overlooked aspect of audio-language understanding and provides a diagnostic foundation for building future LALMs that are robust to real-world acoustic conditions.
Yize Li, Ningyuan Yang, Sile Yin et al.· 0 citations
Humans naturally exhibit multiple forms of abstraction in reasoning and interaction, including temporal abstraction across decision timescales and strategic abstraction over communicative intents. Inspired by these complementary abstractions, we propose a two-level hierarchical reinforcement learning (HRL) framework for conversational agents that bridges the gap between existing token-level and utterance-level RL methods. Built upon a two-level Markov decision process (MDP), our framework conditions token-level response generation on utterance-level actions represented by explicit textual strategies. Based on theoretical analysis and efficiency considerations, we employ DQN to optimize the high-level Q-network and PPO to train the low-level actor-critic. To further alleviate reward sparsity and facilitate convergence, we introduce a dual-granularity reward mechanism that combines the utterance-level satisfaction score with token-level intrinsic self-consistency and a KL-divergence penalty. Experiments on both daily-life and emotional support conversations demonstrate that our method consistently outperforms a wide range of baselines in both strategy determination and response quality. Our implementation is available at https://github.com/AaronJi/ToSCA.
Xiaoyu Wang, Qingqing Gu, Yue Zhao et al.· 0 citations
As lifelong learning agents accumulate lifelong growing skill banks, retrieving the correct skill becomes an increasingly important bottleneck. Most current skill retrieval methods treat each skill as one flat document by concatenating fields such as the name, description, and body. However, skills are naturally structured, multi-field objects, where each field provides different information about when and how the skill should be used. In this work, we study whether preserving this structure improves skill retrieval. We represent each skill as its separate components, and compute sparse and dense similarities for each field independently, exposing a naturally tensorized, field-aware representation of the skill bank. We then combine these field-level scores either with uniform weights or with a small learned MLP. Across two different skill retrieval benchmarks, SkillRet and SRA-Bench, we find that keeping fields separate improves hybrid retrieval, and learning over the field-level scores gives the strongest and most consistent results. Our field-aware MLP reaches $77.95$ Recall@10 on SkillRet and $83.78$ Recall@10 on SRA-Bench, outperforming the corresponding concatenated learned baselines. We also find that the advantage grows as the skill bank becomes larger, suggesting that field-aware skill retrieval becomes especially useful in the setting where retrieval is most difficult. Our results show that skill representation itself matters, and that simply preserving the structure already present in skill files can substantially improve retrieval.
Paimon Goulart, Liang Wu, Kelly Wan et al.· 0 citations
Echocardiography is the most widely used non-invasive cardiac imaging modality, providing essential information for cardiovascular diagnosis. Interpreting an echocardiogram requires synthesizing complementary evidence across multiple heart views to identify abnormalities and produce structured clinical reports. While recent efforts focus on improving classification performance, most models lack explicit diagnostic reasoning and spatially grounded anatomical evidence, limiting clinician trust. We present EchoSonar-R, a multi-view reasoning-enabled vision-language model that jointly performs multi-label disease classification and report generation from echocardiography studies. EchoSonar-R combines a spatiotemporal video encoder with a structure-aware cardiac detector that provides spatially grounded anatomical cues to improve interpretability and clinician trust during cross-view reasoning. EchoSonar-R is trained in two stages: supervised fine-tuning (SFT) on reasoning-annotated targets, followed by Group Relative Policy Optimization (GRPO) with task-specific rewards that jointly align classification and report generation within a unified reinforcement-learning framework. Across a private multi-view dataset and two public benchmarks, EchoSonar-R improves macro balanced accuracy by 17.1% on the private set and 6.1% on MIMICEchoQA over the strongest baseline, achieves a GREEN clinical faithfulness score of 0.800, and produces interpretable reasoning traces grounded in multi-view visual evidence.
Darya Taratynova, Ahmed Aly, Numan Saeed et al.· 0 citations
We investigate a structure-first approach to dynamical learning in which the organization of stateful interactions is prescribed explicitly rather than left entirely to a generic recurrent parameterization. We introduce causal recurrent units built from an ordered sequence of local, state-modulated transformations. The construction is motivated by wave-based interaction models, but the units studied here do not impose scattering, passivity, or energy-balance constraints.
Because fixed recurrent dynamics, designed reservoir topologies, readout-only learning, and recurrent depth are already well established, the empirical question is deliberately narrower: does the proposed interaction organization provide a useful inductive bias under controlled computational conditions? We compare a one-layer structured model, a two-layer structured model, and a generic echo-state network (ESN), all with 12 recurrent states and the same strictly linear ridge readout. Each model family receives the same random-search budget on calibration data that are disjoint from the final test data, after which the selected hyperparameters are frozen.
On a custom nonlinear identification task, the one-layer structured model attains a mean validation NMSE of 2.76 x 10^-4, compared with 3.19 x 10^-4 for the two-layer model and 3.94 x 10^-4 for the ESN. On NARMA10 the ordering reverses: the ESN attains 0.312, compared with 0.348 and 0.357 for the one- and two-layer structured models. Thus, the proposed organization can be competitive and advantageous on one task, but it is not universally superior; moreover, recurrent depth does not provide a systematic benefit under matched state dimension. The results support a task-dependent interpretation of structural inductive bias and position the present architecture as a controlled precursor to stronger wave- and system-theoretic constructions.
Long-horizon agents rely on memory mechanisms to compress interaction history, but optimizing memory writing faces a distinct credit assignment challenge: a memory update may be rewarded or penalized due to downstream tool failures, noisy observations, or reasoning errors rather than its own contribution. We propose HiMPO, a Hindsight-Informed Memory Policy Optimization framework for assigning less-entangled credit to memory-writing actions in long-horizon agents. HiMPO first estimates the local utility of a memory update by comparing the task-relevant information recoverable from the previous and updated memories under the same pre-write state. It then uses hindsight relevance as a bounded retrospective filter that attenuates memory credit when local utility is not supported by the target outcome. The resulting memory-specific advantage is applied only to memory tokens, while trajectory-level rewards optimize the rest of the agent's behavior. Across judge-based open-domain tasks and objective compressive-memory QA, HiMPO improves over strong memory-based and RL-based baselines while preserving compressed-context efficiency. Controlled interventions and live replay studies further show that HiMPO reduces blame leakage from tool-induced errors, assigns memory credit that aligns with the functional impact of memory writes, and remains robust to noisy training targets.
Jiangze Yan, Yi Shen, Wenjing Zhang et al.· 0 citations
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
MIT News · Artificial Intelligence· news.mit.eduAug 24, 2026
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
