Papers

Large language model (LLM) assistants can now search the web and consult external sources while completing user requests. These sources can provide useful evidence, but they can also introduce additional content into the model's context. Can such passive exposure steer a decision even when the added content provides no reason to change it? We examine the stability of model decisions on the same tasks with and without such external content. Across all open-weight and closed-weight models we test, exposure systematically shifts decisions, with effects reaching nearly 50 percentage points in closed-weight models. The same pattern appears with real-world online opinions. The influence also extends beyond subjective preferences. Such exposure can steer models toward choices that violate explicit user requirements and increase their acceptance of false claims. In short, what enters an LLM's context can influence its decision even when it should not determine it.

Modern LLM evaluation assumes that pinning a judge to a fixed model version and decoding at temperature zero yields reproducible verdicts. We show this assumption fails as a property of how LLM-as-Judge is operationalized on cloud serving infrastructure, not of any particular model family. Across four frontier judges served via a single major enterprise cloud platform and three standard benchmarks (Arena-Hard, AlpacaEval 2, MT-Bench), identical inputs to the same temperature-zero judge, at a constant serving-reported model version, produce different verdicts across re-runs: per-item flip rates of roughly 5% on average and about 40% on the close-call items that decide leaderboard margins, with a per-judge magnitude spanning a 40x range (0.13% to nearly 10%). We introduce metrics tailored to this instability: per-item flip rate, a two-part stability profile (waver fraction and conditional intensity), and adjacency separability. For the principal judge the aggregate ranking is stable (0% top-K instability, 0% pooled winner flip); what degrades is precision: under a paired hierarchical bootstrap, roughly one-fifth to three-quarters of adjacent leaderboard positions are statistically indistinguishable, a noise floor driven mainly by finite prompt sampling rather than the judge. Across judges, leaderboards agree on the coarse ordering but diverge in the middle (Kendall's tau of 0.42-0.64 between Gemini and Sonnet judges on Arena-Hard, values sensitive to answers truncated at the generation cap). Of 15 expected head-to-head orderings we re-judge, 12 survive every re-run of the principal judge but only 8 survive every judge. Leaderboards thus report unhedged point estimates that overstate their precision. We propose a minimal, low-cost reporting protocol: several judge re-runs, published stability profiles and adjacency intervals, and results under at least two judges from different families.

We propose DAST (Diversifying Arithmetic Sampling with TokenTour), a method that increases the diversity of LLM outputs without any change to the marginal distribution and with negligible generation-time overhead (a few microseconds). We observe that token IDs are often arranged in a meaningless order and reassign them so that tokens with similar meanings appear consecutively. This can be done in advance in a few hundred seconds per model, and the resulting order can be reused for all subsequent generations. By combining this order with arithmetic sampling (or quasi-Monte Carlo methods), we make similar tokens less likely to be generated across runs while preserving the distribution. Our method not only produces qualitatively good ideas but also significantly improves performance on the downstream task of ProtoQA.

Post-training large language models (LLMs) on two-player zero-sum games provides a verifiable reward, and the gains have been shown to transfer to general reasoning. Existing pipelines, however, rely on a handful of hand-written games and on advantage estimators tied to them, so both the supply of games and the advantage estimate on a large corpus limit scaling. We present GameSmith, which scales both. An LLM synthesises games inside templates whose reward, observation, and hidden-information code is fixed. Each game is verified by seed-paired replay and by an independent re-implementation of its specification, then expanded into encoding, rule, and size variants, giving 26 new games in 29 families and 336 verified variants from the three SPIRAL templates. The model is distilled from the self-play of a larger teacher (Qwen3-32B) on these games and then trained by self-play reinforcement learning with CRAB, an advantage estimator that credits each decision against a shrinkage estimate of the expected return of its position and is defined on every variant. On Qwen3-4B-Base, distillation raises the 13-benchmark mean pass@1 from 45.5 to 62.5, above the three SPIRAL games at the same budget, and self-play with CRAB reaches 64.9. The ordering holds across models, and transfer grows with the number of game families.

Evaluating recommender systems before deployment requires simulators that remain realistic over long, closed-loop interaction cycles. Current LLM-based agents generally function as surface imitators that generate realistic short-term responses but fail to capture the underlying user dynamics such as fatigue, affect, and shifting preferences. We propose PsyBer-Agent, a psychology-inspired simulator that models these dynamics through its Psy-Engine. This engine represents each user with three evolving, model-defined latent states. To calibrate simulated behavior against real interaction logs, we align log-conditioned reference trajectories with simulator-generated latent-state trajectories via Unbalanced Gromov–Wasserstein optimal transport. This approach provides structured supervision for a lightweight policy without the computational burden of fine-tuning the backbone LLM. We also introduce WebSim and the PsyBer Benchmark, which encompass multimodal scenarios and seven recommender backbones. Our framework evaluates behavioral fidelity, latent-state consistency, robustness, and responses to social-cue perturbations. On the tested horizons, PsyBer-Agent shows lower behavioral discrepancy than the prompt-only simulators while maintaining stable performance under large-scale concurrent simulation for up to 5M queued simulation tasks under bounded concurrency. Code is available in the supplementary material and at: PsyBer-Agent: https://anonymous.4open.science/r/CC66F/, WebSim Platform: https://anonymous.4open.science/r/D8EAX/.

Sequence models with a fixed-size state avoid the quadratic cost of attention, but face a design choice. An input-dependent state transition gives content-dependent memory at the cost of FFT-based training, while a fixed transition keeps the O(T log T) convolutional path but reads memory the same way for every input. ResonatorLM, a recent model of the second kind, reads a bank of damped complex oscillators through a fixed learned phase per head. We make that readout phase a function of the current token, which adds input dependence to the read only, keeps the convolution exact, and costs 0.2 percent more parameters. An analysis of the untrained bank shows that phase does encode elapsed time, but an ordinary token falls below the background within 2-8 steps and no choice of damping extends linear reconstruction beyond about 16 tokens. In 6M-parameter character-level models on WikiText-2, the input-dependent phase matches the original at the training length and removes its failure beyond it: trained on 256 tokens and tested on 1024, perplexity falls from 12.81 to 3.76. We trace the failure to long-memory heads that accumulate more history than training exposed them to. The learned phase tracks the current character rather than distance and does not improve retrieval.

Small language models trained on a single local machine leave little room for architectural trial and error. Kairos combines bidirectional gated-DeltaNet linear attention and sliding-window attention over shared projections (LiZAttention2), block-wise Attention-Residual aggregation, a sparse Mixture-of-Experts (MoE) feed-forward layer and a multi-scale byte-level codec, trained with a uniform-noise denoising objective over one byte space shared by text, image, audio, video, lidar and control streams; this report asks whether the combination optimizes at all before asking whether it is useful. We give code-derived algorithms for each component, including a cross-session recurrent-state memory gate, then run convergence sanity checks: overfittinga single 440-byte sentence at 1–15M parameters, text only, with three seeds for codec and attention choices and single-run sweeps for seven further factors. The two attention variants tested at 3\,000 steps (LiZAttention2 and window attention alone) reach a minimum loss near $10^{-3}$. After 600 steps, the mean minimum loss ranges from 0.018 (patch codec) to 0.354 (depthwise convolutional codec) with an identical ranking on three seeds. An apparent advantage of shared over separate attention projections closes with more steps or more diverse data and reverses when the projections are frozen: it is a convergence-speed effect, not a capacity effect. These checks suggest that the implementation optimizes; they say nothing about generalization, multimodal training or comparison with existing models, which a planned pretraining run is meant to address. The Discussion sizes a 200M-total/25M-active target analytically; none of its throughput figures is measured. Code is available at https://github.com/fabienfrfr/Kairos . Python package at https://pypi.org/project/kairos-fm/ .

Small-data adaptation can improve speech detection while degrading speaker attribution. We study this discrepancy in a released streaming diarizer adapted on 7.5 h of two-party conversation and evaluated across six corpora. Adaptation substantially improves in-domain diarization performance and transfers to an independent corpus. However, this improvement is not consistent across evaluation scenarios as the additional confusion is mainly associated with impaired temporal identity consistency rather than speaker-count errors. A local-remapping diagnostic reveals different patterns of identity degradation across corpora, indicating that adaptation may alter how streaming models maintain speaker assignments over time. Rehearsal reduces the observed degradation but reduces the cross-domain transfer performance. These results highlight the need to jointly evaluate detection accuracy, identity consistency, and retention behavior when adapting streaming diarization systems.

Token-level certainty is widely used as a proxy for correctness in LLM training and inference. However, the performance of certainty-based methods depends both on the information in certainty scores and on how those scores are used. We therefore directly assess certainty's predictive ability through controlled empirical evaluations across models and tasks. We distinguish two prediction targets: identifying questions a model is more likely to answer correctly and distinguishing correct from incorrect responses to the same question. In our experiments, certainty is generally better at identifying questions a model is likely to answer correctly than at distinguishing correct from incorrect responses to the same question. Certainty also varies systematically across token types and positions within words, reflecting local properties of words and text form. Information about question difficulty appears early in generation, while the weaker information about answer correctness is more concentrated near the end. These findings show that the information certainty provides for decisions depends on the prediction target, the model, the certainty metric, and which token positions in the response are included in aggregation. We further demonstrate the practical value of these findings for test-time compute. We allocate the number of responses using certainty early in generation and weight answer votes using certainty near the end of each response. Compared with a fixed-sampling majority-voting baseline, this approach increases overall accuracy from 78.71% to 79.54% while reducing generated-token cost by 82.4%.

Benchmarks for large language models (LLMs) typically evaluate the accuracy of answers against fixed reference labels. But a central step in many complex real-world tasks is identifying which questions are worth asking in the first place: decomposing a difficult problem into subquestions – which we call cruxes – whose answers provide key steps on the path toward solving the target problem. To evaluate this capability of information discovery, we introduce CruxBench, a benchmark that grades LLM-generated questions by their Value of Information (VOI): how much a model-proposed crux updates beliefs about a target forecasting question. CruxBench enjoys a rare combination of three key properties: it is (1) contamination-resistant by construction, since ground truth is generated by future world events; (2) open-ended, admitting unbounded and complex text-based submissions rather than one correct numeric answer; and (3) grounded, with informativeness measured against quantified changes in real-world beliefs. We evaluate a diverse set of eight models on 293 target forecasting questions and find that VOI correlates highly with independent measures of model capability (r=0.90) and captures cruxes' usefulness for answering target questions. However, information discovery remains challenging even for frontier LLMs, which only narrowly outperform a random-timing baseline.

Pre-tokenisation restricts which text fragments can become prediction units, but its compression cost is obscured when tokenisers are compared only under the same boundaries. We measure this cost by bounding the minimum token count from both sides, with and without a regular-expression boundary rule. Nonnegative prices on token occurrences yield a lower bound through shortest paths and vocabulary-budget selection; maximising over all prices recovers the linear programming relaxation, and an independent integer checker certifies the reported values. On English Wikipedia, boundaries increase the optimal token count by 28.3–36.8%. Byte pair encoding lies 2.1% above the constrained lower bound, but 10.9% above the unrestricted bound. Compression and prediction favour different dictionaries: at 85M non-embedding parameters and matched training-token budgets, unrestricted fitting yields higher mean held-out bits per byte under a common unrestricted decoder in all 12 languages in the paired study and 11 of 12 under independent tuning and evaluation. To study intermediate boundary policies, we introduce boundary licences, which limit the vocabulary entries permitted to cross cuts and admit the same form of certificate. On separate English and Chinese fitting corpora, licensing 10% of the vocabulary budget recovers 85.2% and 100.0% of the achieved token-count reduction from removing all cuts. These results quantify the compression cost of boundaries while separating it from the prediction quality of the resulting token units.

Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs—two contexts that differ in one edited fact that flips the answer—each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy ($84.6\%$ vs. $85.2\%$; McNemar $p \ge 0.50$ at every seed) while giving the lowest position bias of all runs (answer flips under relabelling $9.8\%$ vs. $13.8\%$) and the lowest ordinal error on rubric fields (MAE $0.232$ vs. $0.311$). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by $26\%$. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions.

Action agents do things for people. They send email, spend money, and call businesses while the user is busy with something else, so a mistake can turn into an action before anyone notices. They fail their users in two ways. They break trust when they do something the user never agreed to, or hold back after the user clearly said go. And they fall short on completion when they give up on errands that turn out to be hard. Both depend heavily on the harness around the model, meaning its instructions, tools, context, and guardrails. We built an evaluation that scores trust and completion on the same runs, in a simulated world of businesses with their own websites, inboxes, and phone lines, and of people who write back. A simulated user answers the assistant's questions. Trust means that nothing happens the user did not agree to. No email goes to someone they never approved, no private detail ends up on a group thread, no money is spent past their limit, no stranger's instructions are followed, and nothing is claimed without a source. Every trap has a matched control in which acting is the right call. We use the evaluation to measure Fo, Wajo's personal assistant, against a base model with basic instructions on three foundation models, and against the Fo harness with its guardrails switched off. Fo completes 71% of the errands and keeps the user's trust on 94% of the trap runs. The base models complete 50% to 64% and keep trust on 59% to 75%. On the matched controls, Fo goes ahead slightly less often. OpenClaw, a popular open-source assistant given the same access, completes 42% of the errands it shares with Fo, against 71%, and keeps the user's trust on 74% of the shared trap runs, against 94%. Measuring trust and completion together, on the whole system rather than the model alone, is how we think action agents become safe to hand real work to.

Medical benchmarks for language models are built almost entirely on Western biomedicine. Traditional Chinese Medicine (TCM) is a separate system, with its own diagnostic framework and its own literature, and it remains largely unmeasured. The few TCM evaluations that exist are small, narrow, and rarely paired with a human reference. We present TCMQA, an open benchmark of 38,279 questions from Chinese TCM licensing examinations, paired with 15,151 responses from 101 licensed practitioners. We evaluate 29 instruction-tuned models from 9 families, spanning 0.27B to 14.8B parameters. Accuracy ranges over 59 points, and no model approaches saturation. Pretraining data predicts TCM ability far better than scale: a 12B Western-pretrained model reaches 39.6%, while a Chinese-pretrained model an eighth its size reaches 60.8%. Nine models exceed the practitioner majority vote of 64.9%, the best by 21.8 points, and all nine come from that same Chinese-pretrained family. Yet difficulty does not transfer between models and practitioners: accuracy is flat across practitioner-rated difficulty, item-level agreement is near zero for all 29 models, and on $8.4\%$ of items the practitioners are correct where the leading model is wrong. We release the corpus, the practitioner responses, the harness, and per-item model outputs at https://huggingface.co/datasets/TechTCM/TCMQA.

Diagnosing a patient from video requires more than recognizing the sign: a vision-language model must turn what it sees into hypotheses, questions and tests. DynamicDx evaluates each step in 71 neurological consultations across 11 sign categories, linking authentic patient videos to confirmed diagnoses and fixed charts built from the same case reports, so that every model queries the same evidence. Across five such models, video improves accuracy by 9.9-22.5 percentage points over blind input, but neither recognition alone nor temporal order explains the gain: the cause is usually missing from the model's video-only differential diagnosis even when the sign is recognized, and shuffling the frames produces no reliable accuracy loss. Instead, a trajectory replay traces most of the gain to the investigation results the video prompts. Evidence acquisition is the bottleneck: supplying the decisive investigations raises accuracy to 73.2-93.0%. Two interventions act on it. A post-trained 4B video describer improves sign descriptions, especially from a short, densely sampled segment, and source-clean literature retrieval expands initial hypotheses; both bring the tests a model orders closer to those the treating clinicians documented and, through them, raise accuracy. For video-based diagnosis, seeing better helps when it leads to asking better.

A two-hop question makes a language model retrieve twice: once to produce a bridge entity, and once to retrieve with it. Transformers resolve that entity in their early layers. Hybrid models replace most of the attention with a recurrent state, so where the key becomes usable, and where it is spent on an answer, is not known. Answering "Where is the ball?" from "the ball belongs to Alice" and "Alice is in the garden" turns on Alice, a name the question does not mention. A model could reach garden through Alice, the key it computed, or through where the fact sits in the prompt. Across twelve models, dense and hybrid, we move a hidden state from one story into another where the two routes lead to different places, and read off which place the model gives. An attention layer converts the key in every model we tested, and crossing that layer removes its usable effect, all of it where no attention follows. In sequential hybrids this makes the answer a handoff: recurrent layers carry the key forward, and attention spends it. Retrieval does not always stop there. Writing a different fact into the memory of the recurrent layers that follow the last attention layer can move the answer toward that fact, multiplying its odds by 1.3 to 2.7, and which hybrids do this is not settled by their architecture or training. That read is addressed by the key, not by position: recurrent state in a hybrid is not only a carrier, but a memory that later layers can query.

Multi-agent systems built on large language models (LLMs) are largely homogeneous, as their agents behave alike even across distinct LLMs. We show that when such agents act concurrently without communication, they collide on targets they must split and diverge on targets they must take together, a double failure we term the symmetry trap. Theory of Mind (ToM), widely used for coordination without communication, cannot escape this trap, since homogeneous agents form the same prediction of one another and respond to it in the same way. We propose Theory of Scene (ToS), a training-free reasoning schema in which each agent reads its public role, the only difference between the agents, and the task context they all observe. Homogeneous agents thereby derive one division of labor, each taking the part its role fixes, which turns homogeneity from the cause of the trap into the cure. ToS reads the role together with the scene through role gating, which determines whether ownership overlaps or is already divided, and the task context through task coupling, which infers whether the team must converge on each target, divide it, or take its stages in turn. We evaluate on DivvyBench, a controlled environment we introduce, whose target types make an episode Competitive, Cooperative, or Mixed across Tabletop, Airspace, and Household scenarios, and on two established agentic benchmarks, GovSim and Overcooked. ToS outperforms all six baselines on every benchmark, and each baseline falls far behind it in at least one setting. Against ToM given the same inputs, ToS raises the DivvyBench success rate from 71.1% to 99.6%, the GovSim total gain from 207 to 400, and the Overcooked level-normalized throughput from 1.41 to 1.67.

Running large language model (LLM) agents in production gets expensive fast. A frontier model (the largest, most capable tier) is accurate but can cost 25 times what a small model costs per token, and the gap compounds once a workflow chains several calls together. Planner-as-Router (PaR) attacks this from a different angle. Instead of leaving model-tier selection to some component downstream, it folds the choice into planning itself. As the planner breaks a query into subtasks, it also assigns each one a model size tier (small, mid, or frontier, ordered by capability and price), so the dependencies between subtasks are visible before any specialist runs. Unlike per-call routers such as cascade routing, which look at one node at a time, PaR sees the whole workflow up front and needs no separate router model or training data. We evaluate PaR with EntBench, a benchmark of 54 enterprise agentic tasks across seven classes, graded by actually running the generated Structured Query Language (SQL) and MongoDB queries against live databases. Over 1,157 evaluations spanning eight routers and three seeds, PaR stays on the observed cost-accuracy frontier. It matches a sink-frontier heuristic (frontier model on terminal nodes only) in accuracy at comparable cost and a faithful FrugalGPT cascade at lower cost, and cuts cost 44% against all-frontier routing while giving up 2.9 points of accuracy. Several accuracy gaps fall inside the plus-or-minus six-point confidence interval of a 54-task study, so we frame PaR's advantage as frontier position rather than a clean accuracy win. We also report a preliminary observation, not a validated result: a small pilot hints that cheap routing may carry a hidden compounding penalty on compositional workflows, which we frame as a hypothesis for future measurement. PaR, EntBench, and all evaluation code are open source.

Recent work has identified a human brain network specialized for abstract formal reasoning (Kean et al., 2025). Does the same hold true in language models? To answer this question, we introduce the minimal viable subspace (MVS) method, which searches for the lowest-rank activation subspace at a layer that preserves task performance when everything outside that subspace is ablated. Using MVS, we demonstrate low-rank subspaces supporting logical inference on Gemma and Qwen models. Furthermore, these subspaces exhibit a clear dissociation from model capacities on other tasks, such that retaining these late logic subspaces preserves inference while impairing factual knowledge, working memory, cognitive control, and arithmetic. Conversely, ablating them reduces logical inference accuracy to chance while largely sparing these other capacities. Our results suggest a functionally localizable core machinery for logic akin to that in the human brain.

Training language models with ternary weights is commonly judged by loss and downstream accuracy, which record only a modest cost relative to full precision. We show that these metrics can conceal a much larger failure. We instead measure knowledge capacity, the factual bits stored per parameter, on synthetic biographies with known information content. We train GPT-2-style models from scratch with 2.5M to 50M parameters at five precisions. Under the standard cosine schedule, ternary models retain as little as 6% of an identically trained fp16 model's capacity. The deficit widens with model size while perplexity rises by only 1.4 to 1.6 times. Measured throughout training, these models first acquire capacity and then lose most of it. We identify this knowledge collapse as a learning-rate dwell instability. Held near $1$–$2\times10^{-4}$ with no decay, a pre-collapse model collapses within a few hundred exposures, and returning to a safe rate does not restore capacity. We then locate the collapse in the output head. It happens at a value the model can never predict. The weights there grow unchecked, while every other prediction the model makes is unchanged. We find that making the value predictable removes the collapse, regardless of which attribute carries it. A warmup-stable-decay schedule with a 10% cooldown increases ternary capacity by 2.7 times at 25M and 4.3 times at 50M. Gains are larger at lower precision. Cutting the output head's learning rate prevents failure when training from scratch. The post-training quantization methods we tested recover no measurable capacity below 4 bits. Our findings suggest judging low-precision training by retained capacity during training rather than final loss.

State-of-the-art pathology foundation models, trained on millions of histology tiles, can fail to preserve tissue similarity when comparisons cross slide or institution boundaries. We show that general-purpose multimodal LLMs, without being trained as pathology foundation models, consistently outperform these specialized models in cross-domain histological similarity judgments. Using a relative similarity framework that we release as the MOSAIC (Model Similarity Assessment across Institutions and Cohorts) benchmark, we evaluate 17 models across 6 datasets and find that pathology encoders often rank same-institution, different-disease tiles as more similar than same-disease, different-institution tiles, a clinically dangerous failure mode invisible to standard within-domain evaluations. LLMs appear less susceptible to this failure, likely because they perform semantic visual comparison of morphology and tissue architecture rather than relying on shortcut features tied to acquisition context. Scaling training data does not resolve the problem for pathology encoders, implicating the learning objective rather than data coverage. Our results expose a fundamental robustness gap in current pathology foundation models and establish multimodal LLMs as a viable alternative for cross-institutional retrieval, dataset harmonization, and multi-site quality control. Code and data will be released upon acceptance.

Sycophancy is a language model's tendency to cave when a user pushes back, abandoning a correct answer for the user's. Several benchmarks now measure it by scripting an objection and recording how often the model caves. Because that objection is a prompt template, whatever else the template varies is measured along with the property it claims to isolate. We audit SycEval, which reports that objections raised before a model answers (preemptive) cause more caving than those raised after (in-context), and attributes the gap to timing. Two features of its templates vary alongside the property each is meant to test. First, SycEval's objections escalate through four strength levels, and at the two weakest only the preemptive template names a target answer, so timing and naming vary together. We build the missing comparison and test it on multiple-choice questions and SycEval's own free-form pipeline. Naming a target answer raises the follow rate, the share of samples matching the user's assertion, by 14.1–49.5 percentage points (pp), and once both templates name one, the timing comparison reverses on three of the five model conditions we test: models cave less under preemptive objections than in-context ones, opposite to SycEval. Second, the output-format instruction benchmarks append for automatic grading also varies with objection placement. Moving it from the pushback into the question flips its effect in opposite directions across models on the same items ($p=0.0059$). The same effect appears in SycEval's free-form pipeline: relocating its instruction lowers caving by 5.1pp on Llama-3.1-8B, while an equivalence test confirms no effect on Qwen3-4B. Both confounds live in the template rather than the models under test, so both are correctable: we close with three checks benchmark authors can apply before publishing.

While Large Language Model (LLM)-based agents demonstrate strong capabilities in long-horizon tasks by interleaving reasoning with external environment interactions, the continuous accumulation of context rapidly creates a critical memory bottleneck. Existing memory compression methods rely on task-specific optimization or external auxiliary models, introducing significant computational overhead. Furthermore, the resulting compressed representations tend to lose structured relationships, leading to information dilution, attention collapse, and degraded decision consistency. To address these limitations, we propose Auto-Regressive State Machine (ARSM), a lightweight training-free framework that enables in-situ reasoning compression through structured state evolution. ARSM introduces two key components: (i) a trajectory abstraction mechanism that reorganizes interaction histories into compact Hypothesis-Action-Result (HAR) micro-chains; (ii) a dynamic state machine that regulates hierarchical memory through atomic operations and a compression-control parameter. These components are unified within an auto-regressive, self-compressive generation space, where each model output jointly performs external action execution and internal state updates. We evaluate ARSM on Webshop, Multi-Objective Multi-Hop QA, and SWE-Bench Lite datasets. Experimental results show that ARSM maintains the task performance while simultaneously reducing token consumption, offering a practical, cost-effective route toward scalable autonomous agents for long-horizon tasks.

A four-stage LLM pipeline gives up as much as 40.5 accuracy points at its own interfaces (gemma-3-12B on MATH-500, Holm-corrected p = 1.66e-19; the largest tax in the primary family). We hold model, problem, stages, stage prompts and completion budget fixed, vary only whether each stage can still see the original problem, and call the accuracy difference the decomposition tax. Across 21 open-weight models from nine organisations, on GSM-Hard and MATH-500 at n = 200 paired items per cell, 70 of 118 primary-family tests survive Benjamini-Hochberg correction and 54 survive Holm. On GSM-Hard, a placebo recovers nothing: it carries at least 60% of the extra tokens and at most one word of the problem. Builders design a pipeline one stage at a time, and its bill arrives at the interfaces between stages. Rewriting one stage's instruction moves gemma-3-12B's tax from 4.5 to 36.5 points, and adding "every relationship stated between them" to a stage that lists the numerical quantities lowers the tax on 9 of 9 models on MATH-500. Re-grounding, which shows a stage the original problem again, belongs after the loss. With one lossy interface, re-grounding the stage after it beats re-grounding the stage before it on 7 of 7 models on both benchmarks; on MATH-500 the earlier repair is worse than none on 7 of 7. Newer models still pay: gemma-4-12B gives up 37.0 points, and the repair holds on all three of the newest models we test. A sealed held-out test refuted a stronger rule we registered, which predicted the paying stage from the interface and receiver types, so we locate the tax by measuring one stage at a time. The prescription has two parts: re-ground the stage after the lossy interface, and if a stage must list the quantities, tell it to keep the relationships.

Large vision-language models (VLMs) have emerged as a powerful paradigm for urban and spatial AI. However, current state-of-the-art large VLMs still struggle with quantitative reasoning on remote sensing imagery. Existing benchmarks and algorithms are predominantly based on qualitative Visual Question Answering (VQA), providing limited insights into the quantitative reasoning capabilities of VLMs for built environment metrics. To address this gap, we develop Quantitative Urban and Spatial AI benchmark (USAI-Quant), the first benchmark designed to quantitatively evaluate VLM's reasoning capabilities on built environment metrics via remote sensing imagery. USAI-Quant is curated from the 335 largest U.S. cities, aligning high-resolution remote sensing images with quantitative built environment metrics. We then evaluate both general-purpose and remote sensing VLMs (RS-VLMs) by applying VQAs to tens of built environment metrics across three complexity levels. Our results reveal that current state-of-the-art models consistently fall short on numeric reasoning tasks. We further conduct in-depth analyses across models, question types, and geographic locations, uncovering insights into performance variability and task-specific challenges.

Multidisciplinary tumor boards integrate multimodal clinical observations and longitudinal patient histories through specialist discussions, yet benchmarks rarely capture these real-world trajectories. We introduce OpenTumorBoard, a benchmark with 611 patient cases and 19,157 discussion turns across ten specialist roles, transcribed from 12,534 minutes of publicly available tumor board recordings on YouTube. The benchmark evaluates two settings: SPECIALIST TURN, in which an LLM responds to a clinically significant question posed during a real discussion, and BOARD SIMULATION, in which it generates an entire back-and-forth discussion and reaches a consensus on therapy recommendations, surgical plans, next actions and clinical trial matching. Evaluation of 14 general-purpose frontier and medical LLMs reveals substantial limitations: the best models score 3.43 out of 5 in clinical equivalence to specialist answers and 2.78 out of 5 in alignment with recorded board conclusions. Supervised finetuning and reinforcement learning improve performance on a held-out test set, suggesting that real-world discussion trajectories can support model adaptation. Three M.D. experts review a subset of the benchmark, finding high information coverage and factuality of patient cases and strong fidelity of extracted consensus conclusions. We will release OpenTumorBoard and its automated curation pipeline to support the development and evaluation of LLMs for multidisciplinary, personalized cancer decision-making.

Multiple-choice and candidate-selection evaluations are widely used to assess LLM reasoning and decision-making, yet most benchmarks contain relatively small candidate sets. It remains unclear whether conclusions drawn from these settings remain valid as the candidate space scales. We systematically evaluate LLMs as the number of competing candidates increases and find substantial accuracy degradation across tasks, prompting strategies, and model scales. Controlled analyses show that standard long-context retrieval explanations cannot fully account for this degradation. Instead, we identify two systematic failure patterns. First, gold-margin collapse: the score gap between the correct answer and the strongest distractor progressively shrinks, driven primarily by weakening confidence in the correct answer. Second, earlier candidate preferences become increasingly difficult to overturn, with later candidates exerting progressively weaker influence on the final prediction. Motivated by these findings, we evaluate hierarchical partitioning and permutation-based inference, which improve accuracy by roughly 20 percentage points at $N=160$ on both HotpotQA and MIMIC. Overall, our results identify candidate-set scale as an important evaluation-protocol variable and show that strong small-option performance does not necessarily imply robust large-scale candidate comparison.

Large vision-language models (LVLMs) inherit massive activations from their text-only bases: spikes where a few fixed hidden channels receive values thousands of times above the typical magnitude. The text spike systematically appears in early layers at a fixed initial position, independently of input content. Visual spikes vary across images, but whether their formation follows a consistent pattern across LVLMs and how they respond to image perturbations remain open questions. We find that some LVLMs do not form visual spikes, while others spike at different rates, typically in deeper layers. We identify the trigger direction from model weights and an interpretable location rule: before the language model decoder runs, eventual spike tokens are largely restricted to those sharing least with the rest of the image. Crucially, visual spikes are strikingly brittle. Common corruptions frequently create and relocate spikes, and less often remove them, raising overall incidence. Our trigger-guided spike attack deliberately creates or removes spikes under a small $\ell_\infty$ budget, with 1/255 enough in nine of the ten models that spike. Finally, our preventive intervention removes only the trigger component before spikes erupt, eliminating or substantially reducing spikes on clean and perturbed images while leaving the other image tokens nearly unchanged. Our study spans 25 adapter-based LVLMs built on 18 released text-only bases from 10 families, ranging from 2B to 72B parameters.

Long tasks produce more history than an LLM agent can hold in its context, and more than it uses reliably even when the history fits. A growing line of work therefore has agents carry a short written state instead: at every step a writer rewrites the state, and a reader acts from the state alone. Steps stay cheap, but anything the writer drops is lost before later decisions reveal that they need it. We quantify this loss and ask whether training can reduce it. Comparing the written state with the best state of the same size written in hindsight, we split the reader's loss into a budget loss, which any state of that size must incur, and a write-time regret, which comes from the writer's choices. In TextWorld cooking games where we control how long a fact must be carried before it is needed, a 128-token state holding the facts wins nearly every game, while prompted language-model writers win at most 17%. Almost all of the loss is write-time regret, and it grows with the delay. We then train the writer from the reader's own loss. DSSR (decision-sufficient state representations) scores candidate states by how well the reader acts after the writer carries them forward, and teaches the writer to prefer the better ones. This forward-rolled score predicts game outcomes ($\rho = 0.48$), whereas scoring a candidate as a fixed context, as hindsight methods usually do, does not ($\rho \leq 0.07$). On a pre-registered test split opened once, training adds +7.0 [+1.9, +12.2] points of success when facts are needed soon, bringing a plain summary writer to the level of belief- and slot-based memory prompts. The gain shrinks as the delay grows and is significant only at the shortest delay. We trace this limit to credit assignment: keeping a fact now pays off only if every later rewrite keeps it too, which a per-step score cannot see.

One variable sets what an upstream fault costs a staged pipeline of language-model agents: re-derivability, how much of what a stage needs it can rebuild from the original problem. Grounding an inspector agent in that problem is worth +0.608 [+0.517, +0.700] to +0.358 over a blind one on four open-weight backbones served with thinking disabled, and on the two Qwen backbones the blind inspector changes no item at all. That head-to-head is exploratory. One deterministic fault enters the first stage, and we re-expose the original problem to $k = 0,\dots,3$ of the downstream stages with agents, items, fault and topology held fixed, on 120 gsm_hard items per arm at temperature zero. Accuracy under fault rises on four of four backbones, from +0.233 to +0.392, the largest Holm-adjusted $p$ being $2.1\times10^{-6}$. A registered kill test rules out tokens. Blanking every word holds the word slots fixed, and retention tracks the visible fraction on four of four, climbing from 0.221 to 0.692 on the primary. Those two families are confirmatory and everything else here is exploratory. The interaction excludes zero on two of four backbones under the registered pipeline, four of four under a three-stage pipeline, and three of four under full message history, the primary at +0.317. On Llama-3.1-8B the fault carries no detectable cost at any dose, so the other three carry every claim about what a fault costs. Re-derivability also sets what the architecture costs, and no decomposition we measured reliably beats one direct call. With no fault injected the registered pipeline loses to that call by -0.267, -0.125 and -0.317, and on Phi-4 reads +0.058 at $p = 0.118$, which the test fails to separate from zero. The repair that works is cheap and front-loaded: the first re-grounded stage buys +0.394 of matched retention for +59.8 tokens per item on Qwen3-14B, and the stages after it buy nothing.