Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

RAW: Refusal Aversion in the Wild, A Causal Measurement Method for Deployed LLMs

arXiv:2611.21641v1 Announce Type: new Abstract: Deployed LLMs field hundreds of millions of conversations a day and refuse a non-trivial fraction of user requests, yet the user-side cost of refusal, whether users disengage and by how much, is unmeasured on any production model: a blind spot for the consumer-facing platforms built on them. A/B-testing refusal is ethically and operationally constrained on frontier-lab products and unavailable to third-party platforms deploying these models through a fixed API. We introduce RAW, a method that uses LLM stochasticity as a natural experiment: at near-identical prompts in any existing log, the same model sometimes refuses and sometimes does not, and that variation identifies the causal user-side effect of refusal. Applying the method to 15 deployed LLMs in LMSYS-Chat-1M and WildChat, with user-clustered inference, we find that for about half of the models we can measure precisely, a refusal costs 3 to 17 pp of subsequent re-engagement; the rest show no distinguishable effect, among them one deployment whose positive point estimate we trace to persistent users retrying the refused request. Our contribution is the measurement method, which runs on any deployed model from logs alone for about \$0.30 per 1,000 conversations. We release labels, moderation scores, prompt embeddings, and pipeline code.

Regularized Score Centering

arXiv:2611.21791v1 Announce Type: new Abstract: Asynchronous reinforcement learning increases training throughput, but data from stale policies and numerical mismatch between training and inference can destabilize learning. To stabilize off-policy learning, score centering (SC) removes the effect of reward baselines on the expected update. Our gradient diagnostics show less extreme gradients with SC than with importance sampling (IS), but SC can still fail when the learner moves far from the behavior policy that generated the data. To limit this policy mismatch, we introduce Regularized Score Centering (RSC), which adds a reverse Kullback-Leibler (KL) penalty to encourage the learner's next-token distribution to stay close to the behavior policy's next-token distribution. RSC works with or without truncated importance sampling (TIS), which caps token importance weights. Our theoretical analysis supports this approach, showing that KL regularization improves convergence guarantees. We evaluate dense and mixture-of-experts models on mathematical reasoning and agentic tasks, including online shopping, tool use, games, and fully asynchronous software engineering. RSC with TIS substantially improves sample efficiency across all agentic benchmarks, attaining higher training rewards with the same number of sampled responses. In final evaluations, it achieves top performance across all six agentic tasks, improving by 14.3% on AppWorld and reaching 9.9 times the strongest baseline's score on BabyAI.

Internal Verdicts Track Evidence: A Template-Controlled Neutrosophic Reading of Epistemic States in Large Language Models via the Jacobian Lens

arXiv:2611.18602v1 Announce Type: new Abstract: For twenty-five years, neutrosophic logic has assigned independent degrees of truth (T), indeterminacy (I), and falsity (F) to epistemic states that classical and fuzzy semantics cannot express. Previous applications measured these components on outputs: answers, judgments, expert evaluations. This paper reports, to our knowledge, the first neutrosophic reading of components inside the internal representations of a large language model. Using the recently released Jacobian lens on Qwen3.5-4B, we project intermediate-layer readouts onto lexicons of support, refutation, and hedging, obtaining layer-wise (T, I, F) profiles under conflicting evidence, first-person false belief, and factual control (n = 20 each). A first battery yielded three striking signatures: a surge of refutation mass under conflict, sustained indeterminacy under false belief, and an apparent "verdict collapse" before the output layer. A second, template-controlled battery showed that all three, as initially stated, were largely artifacts of prompt grammar: matched templates with agreeing evidence inflate the same lexicon masses, true beliefs elicit the same hedging, and the model verbalizes its verdict when allowed to generate. What survives the controls is stronger than what died: the neutrosophic verdict balance B = log10(F) - log10(T) tracks evidence polarity under identical templates (about 3 orders of magnitude of separation, Cohen's d = 2.3-2.5, p < 10^-6 in 9 of 9 layers; paired Wilcoxon: median difference +2.4 orders, 19/20 items, p = 3.8x10^-6) and residually discriminates false from true first-person beliefs. We argue that the independence of the neutrosophic components is what keeps the common-mode channel observable, making both the measurement and its self-correction possible, and we distill the two-battery design into a reusable protocol for auditing language model internals with lens-based readouts.

VidEngram: Agentic Long-Term Video Understanding via Structured Hierarchical Memory

arXiv:2611.18578v1 Announce Type: new Abstract: Long-video question answering is wasteful and lossy when a multimodal model must re-encode the same video for every query: context limits force sparse sampling or truncation, and no reusable representation remains for later questions. We present VidEngram, an agentic video QA system that builds a persistent hierarchical memory once at ingest and answers later questions from that memory. VidEngram captions video and audio jointly with Qwen2.5-Omni, consolidates segment memories into episode and entity layers, and uses a text-only agentic planner with specialized retrieval tools to produce grounded answers. The web interface makes this process inspectable: users can watch memory formation, inspect retrieval traces, verify cited evidence in the video player, and explore entity relationships. On the full OmniVideoBench, VidEngram with a GPT-5-mini planner outperforms a strong single-pass Qwen2.5-Omni baseline (37.0% vs. 32.3%), with gains concentrated on longer videos where single-pass coverage becomes sparse. Because perception is paid once, subsequent queries use low-cost text-only planning.

Benchmarks Don't Talk Back: Learning Through Self-Disagreement

arXiv:2611.17563v1 Announce Type: new Abstract: A benchmark usually measures a language model alone with a question, and deployment rarely looks like that: answers get contested, and models give up correct ones under counter-reasoning that is itself wrong. We show that what fails there is not a shortfall of accuracy but a capability in its own right -- knowing when to hold a position and when to yield to a better one -- which we call dialectical competence. It stands on first-answer accuracy without following from it: a model can be made materially better at answering and no better at this, and training for accuracy alone can drift against it. Nor can it be bought at inference, by self-reflection or by sampling more. Taking a cue from the dialectical tradition in philosophy, in which productive opposition arises within a position rather than being imposed on it from outside, we hypothesize that this capability can be trained from opposition the model supplies itself. It states a position, generates the strongest case against that position -- its antithesis -- rules on whether the case stands, and is trained on that ruling, with no second model at training or at deployment. We call this learning through Self-Disagreement: reinforcement learning from verifiable rewards throughout, where the same check that scores an answer against its reference scores the ruling too, so no reward model, no preference data and no human annotation enter anywhere. The capability is trained rather than elicited: weak in the base model, it arrives with the reward, is not paid for out of first-answer accuracy, and holds against challengers the model never trained against, including ones stronger than itself. The opposition a model needs in order to improve, it turns out, need not come from outside it.

Can You Trust Your Finance Agent? Drift, Diagnosis and Auditable Intervention on One Shipped Financial Research Agent

arXiv:2611.17939v1 Announce Type: new Abstract: Can you trust a finance agent you did not build? We take the reference agent shipped with the Vals Finance Agent Benchmark v2, leave it untuned, and ask four questions a leaderboard score does not answer. Is the score stable? Seven identical runs on 18 questions pass between 105 and 123 of 149 rubric criteria, and the bill varies by up to 17%. When the agent fails, who owns the fix? We classify 20 failed trajectories, from three runs of nine held-out questions, by the layer whose owner would remedy them, under a taxonomy frozen in advance: 13 belong to the agent's instructions, 3 to its tools, 4 to the model, none to its data. Can that diagnosis be trusted? Four blind readers from three model families, given only the traces, name the same fault as our LLM analyzer in 94% of cases, the rate at which they agree with one another (93%). Can a change be trusted? An optimization loop reads the traces, makes one targeted change per round and keeps it only if it clears a fixed threshold. Under a cost objective its final configuration costs 0.34x the baseline on the held-out questions (95% CI [0.22, 0.58]) at accuracy within noise. Under an accuracy objective it gains 14 points on the development questions and 1.5 on the held-out ones (CI [-5.3, +7.2]). Judged against the measured run-to-run noise, under a rule written before the test was opened, the first is a real saving and the second is not. This is a case study of one agent and nine test questions, not a general result.

Small Open-Weight Language Models for Crowd Decision-Making: Capabilities and Limits of Output-Level Fusion and Linguistic Aggregation

arXiv:2611.17589v1 Announce Type: new Abstract: Small open-weight language models can run crowd decision-making (CDM) on an institution's own infrastructure, which keeps the checkpoint, the inference settings and the opinions under its control. Their judgements vary across repeated queries; whether repeating and aggregating them, often through the 2-tuple linguistic representation, settles the decision rather than the labels is untested. Five open-weight models of 8-36B total parameters on one consumer GPU and a hosted reference classify 474 TripAdvisor reviews of four London restaurants ten times each; three open-weight models repeat the protocol on 600 reviews of twelve Madrid restaurants. Continuous evaluations preserve resolution: snapping each review leaves 5 distinct values against 21-65, and snapping at every level at most two collective values. Sensitivity to the participants persists once run-to-run variability is reduced: in London, redrawing the reviewers keeps the winner on 47-81% of replicas against 86-100% for redrawing the runs. With simultaneous intervals over every pair of alternatives under crowd resampling, no system on either corpus names a single winner: in London every system keeps all four restaurants as candidates, and in Madrid the candidate sets hold five to eight of the twelve. Extra runs and output-level fusion stabilize the winner but do not remove that sensitivity, and the fusion does not beat its best member. The three larger open-weight models name the hosted reference's London winner and at least match its per-review correlation with the reviewers' stars, so small open-weight models can support local CDM, reported as candidate sets with margins rather than forced rankings.

An LLM Is All You Need: Data-Efficient Adaptation to Full-Duplex, Word-Level Spoken Dialogue

arXiv:2611.14774v1 Announce Type: new Abstract: Full-duplex spoken dialogue, in which a system listens while it speaks and can backchannel, yield the floor, and be interrupted, is usually obtained by training on hundreds of thousands to millions of hours of speech. This is costly, rewrites the language model that supplies the system's knowledge, and still leaves turn-taking either implicit or decided coarsely rather than word by word. We show that a frozen, general-purpose multimodal LLM can instead be adapted into a word-level full-duplex system with little data. Our system is a thinker-talker cascade on an 80 ms frame clock. The thinker keeps the LLM frozen and trains only lightweight adapters and a few new token embeddings (97M parameters); at every frame it emits one token of agent text, with pauses, yields, and onsets represented as tokens of the same vocabulary and kept legal by a finite-state machine. A four-stage curriculum, each stage anchored to the model's own behavior under the previous format, performs the conversion in about 30 GPU-hours. The talker reads only the thinker's discrete tokens, so the two halves are trained independently. On Full-Duplex-Bench and InteractionBench, the system is competitive with open-source full-duplex systems built through large-scale speech training: among them, it is the least likely to take the floor when the user merely pauses, backchannels most often, and leads in spoken-task accuracy, groundedness, and coherence over long conversations.

RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving

arXiv:2611.14949v1 Announce Type: new Abstract: Identifying and solving partial differential equations (PDEs) from sampled solution data often requires substantial expert effort in equation discovery, coefficient estimation, and numerical solver configuration, especially under noise and distribution shifts. Can large language model (LLM) agents automate this workflow accurately and efficiently? To investigate this question, we introduce \bench{}, a benchmark of 1,001 cases spanning 42 equation templates, including both classical and fractional PDEs. It provides three evaluation settings: Full evaluates performance on noise-free solution data; Noisy tests robustness to Gaussian noise with an RMS equal to 50\% of the clean-signal RMS; and Hard-OOD tests challenging out-of-distribution generalization to unseen equation templates and shifted coefficient ranges. We further introduce RSI-PDE, a recursively self-improving agent harness that shortlists candidate equations by residual loss, uses relative $L^2$ feedback to refine coefficients and numerical solvers, and applies a validation gate to policy and memory updates. We compare RSI-PDE with single-run baselines and native agent harnesses (Claude Code or Codex) across four frontier models. Across the three settings, RSI-PDE improves the pass rate averaged across the four models by 106.5--181.9\% relative to single-run baselines and by 3.0--19.6\% relative to native harnesses, while reducing total recorded model calls by 46.9\% across the 12 model--setting pairs. These results highlight the potential of validated recursive self-improvement in LLM agents to advance scientific computing.

When Is Proof-Step Verification Worth Its Cost? Operating Requirements of Certify-or-Decline Verification for Open-Weight Mathematical Reasoning

arXiv:2611.14794v1 Announce Type: new Abstract: Proof-step verification rewrites a solver's answer into typed steps and uses LLM judges to certify or decline. We evaluate this pipeline on open-weight models across GPQA, AIME-2025, and a MATH-500 subset. Certification raises answer-key-match precision by 31 percentage points on the strong GPQA configuration and 46 on AIME, while answering only 35% and 21% of attempts. On near-ceiling MATH-500, gains are at most 5 points at a loss of 13-27 points of coverage. An output-token-matched self-consistency selector obtains 82.9% precision at 35% coverage, compared with the verifier's pooled 87.6% (paired difference: +4.8 points, 95% interval [-8.2, 18.5]). A diagnostic sample reveals a distinct failure mode: 52 of 88 completed problems end in invalid proof construction without ever reaching a judge. Finally, external-LLM acceptance of rendered certified chains is 61.0-67.6%, compared with 87.6% answer-key match. These audits lack human or formal-kernel ground truth. Open-weight auditor pairs show near-zero chance-corrected agreement, which measures consistency rather than accuracy against proof truth. These results motivate separate reporting of execution failures, selected-answer precision, and independently assessed proof validity. We provide per-configuration results and evaluation details in the supplement.

INTENT: Out-of-Sample, Out-of-Time, and Cold-Start Evaluation on User Event Sequences

arXiv:2611.13069v1 Announce Type: new Abstract: Banks now train foundation models on ordered sequences of customer events such as transactions, app sessions, and authorizations. Deployed, such a model predicts forward in time, often for customers and merchants absent from training, yet each is evaluated on private data, so no two can be compared. Public benchmarks for such sequences rarely test a period the model has not seen; recommendation studies usually hold out each user's last events instead. We present INTENT (INTEraction, Novelty, and Transfer evaluation for next-action prediction), a public benchmark that scores a model in seven cells defined by whether the user, the period, and the target item appeared in training. An audit of 59 public dataset families retains five datasets; the most common defect is the lack of a calendar timestamp that orders events. No date lies within the time span of all five datasets, so each is split at its own cutoff, chosen by one rule fixed before any model is evaluated. Summed over the datasets, every cell holds at least 111,769 users. In reference results the cell determines the best model: a transformer leads on seen items after the cutoff but scores zero on new items, a variant that embeds item metadata helps mainly on MicroLens, and repeating the user's last item leads on new items in four of five datasets. This parameter-free control also shows that on three datasets most of the score lost after the cutoff reflects changed behavior rather than the model. The pipeline, the cutoff rule, and the per-cell user counts are publicly available.

Measuring Definition Sensitivity in LLM Annotation of Social Science Concepts

arXiv:2611.13339v1 Announce Type: new Abstract: In social science research, text annotation is a fundamental task that is traditionally a costly, hard-to-scale manual process. Researchers have used large language models (LLMs) to perform annotation by writing a definition of a concept into a prompt - a rule that determines whether a label should be attributed to a text depending on its content. However, concepts such as hate speech, fake news, and terrorism typically have several theoretically defensible competing definitions. It is also unknown whether the model actually applied the definition to provide the label or the label was based on the model's pretrained notion of the concept. To test this, we have the model annotate the same texts twice, each time with a different definition. We call the disagreement between the two sets of annotations definition sensitivity. As a control, we paraphrase one definition but leave its requirements unchanged - the disagreement between the annotations with the two wordings becomes the paraphrase floor. Across six concepts and three LLMs, a swap of competing definitions always changes more annotations, on average, than a paraphrase. This shows that the annotations depend on the prompted definition, so any set of annotations should be interpreted together with the definition that produced it. The LLM annotator's definition sensitivity and its paraphrase floor should both be reported in such work.

One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction

arXiv:2611.13001v1 Announce Type: new Abstract: Generation, imputation, and prediction query different aspects of the same tabular distribution, but are usually evaluated separately. This leaves unclear which capabilities they share, how their training objectives interact, and whether one model can perform all three well. We introduce the One Table Benchmark (\ours{}), spanning 56 datasets, 46 registered methods, and 20 capability axes under aligned data splits and explicit task contracts. We compare task ranks on aligned support, train five backbones under all seven non-empty objective combinations on five datasets with three seeds, and contrast fixed-model profiles with specialist portfolios. Rank comparisons connect imputation and prediction through numerical recovery and dependence preservation, and link generation to both through predictive utility; distributional fidelity follows different, sometimes opposing, rankings. Objective interventions show the largest pooled gains from including the recipient's own objective. Generation has the broadest coverage of positive pooled auxiliary effects, while prediction shows potential to improve imputation and synthetic-data utility, with responses that vary across backbones. Cross-addition effects reveal that objective benefits need not add and depend on the objectives already present. Fixed-model comparisons expose corresponding utility--fidelity trade-offs: some generalists surpass specialist portfolios on utility, while fidelity continues to favor specialization. Frozen foundation models show strong prediction and numerical point imputation, together with promising synthetic-data utility, highlighting their potential across all three tasks. These findings motivate unified tabular models that share useful conditional information while improving categorical reconstruction and faithful joint sampling.

PegasusAI Studio: Legible Cyberinfrastructure and Accountable AI Agents for Scientific Workflows

arXiv:2611.13309v1 Announce Type: new Abstract: AI coding agents can now construct scientific workflows, but operating them on shared cyberinfrastructure (CI) raises problems that code generation does not solve. Agents need current knowledge of the workflow system and its execution sites and access to the runtime state of distributed executions, while facilities need to attribute every agent action to a user and bound it by that user's privileges. We present PegasusAI Studio, an AI-assisted science gateway built on the Pegasus and HTCondor stack. Studio makes CI legible to agents by projecting one facility-curated knowledge base into the native configuration of several coding agents, by introspecting workflow interfaces and execution sites, and by grounding diagnosis in workflow and scheduler telemetry. Agent actions operate within a trust envelope that combines federated authentication, per-user execution environments and scheduler credentials, provider-neutral model access, observability, and approval controls for state-changing actions. The same application runs as a JupyterHub-based gateway on a research cloud and as an Open OnDemand application on HPC systems. We evaluate Studio through case studies of legacy CASA weather workflow modernization, a gravitational-wave reanalysis, an EEG analysis built with an open-weight model, and an Event Horizon Telescope (EHT) M87 image reconstruction that recovered from two failed runs to a zero-failure execution. The evaluation establishes that these mechanisms operate as designed.

SELD: Detecting Silent Evidence Loss in Context Compression beyond the Answer

arXiv:2611.12903v1 Announce Type: new Abstract: Context compression is now routine for language models with long inputs, but a compressed context is useful only if it keeps the evidence for the answer. Existing checks compare the reader model's answer or its most likely words, and thus miss a frequent failure, silent evidence loss: key evidence is removed while the answer stays the same, leaving the answer unsupported and risky for other models that reuse the context. To address this challenge, we propose SELD (Silent Evidence Loss Detector), which needs no training and no labels. Its principle is that when evidence is lost, a reader model's preferences among the answers it leaves unsaid shift even if its answer does not, so comparing these preferences before and after compression reveals losses that answer checks cannot see. We also build SEL-Pairs, a benchmark of length-matched evidence and control deletions from three question-answering datasets. Quantitative and qualitative experiments with five reader models on it show that SELD consistently outperforms strong and recent baselines across compressors, compression rates and models, identifying the evidence-removing edit in 83% of matched pairs with unchanged answers, against 70% for the strongest baseline. These results show that evidence loss in compressed contexts can be detected even when the answer is unchanged.

SetPrint: Black-Box LLM Fingerprinting via Setwise Correspondence of Response Changes

arXiv:2611.12950v1 Announce Type: new Abstract: Black-box LLM fingerprinting aims to determine whether a target model was derived from a source model when only the output responses to input queries are observable. One way to reveal this relationship is to perturb each query and observe the resulting responses of both models, yielding two corresponding sets of response variations. These sets provide two complementary sources of information: the relations among response changes within each set, and the correspondence of response changes between the two sets. Existing approaches, however, have not fully exploited both. We introduce \textbf{SetPrint}, a training-free method that jointly captures both within-set relations and between-set correspondence. SetPrint expresses each response change relative to the mean change of its own set, and then compares changes from the two models that were induced by the same perturbation. On LeaFBench, SetPrint consistently outperforms existing training-free black-box fingerprinting methods. Further analyses attribute these gains to both within-set relations and between-set correspondence.

SHADOWPREFILL: Keeping Your Next Model Warm for Per-Step Model Routing in Agentic LLM Serving

arXiv:2611.13176v1 Announce Type: new Abstract: Per-step model routers send each LLM call of an agent to the cost-optimal model dynamically within each session. However, every model switch discards prefix-cache reuse: KV caches are per model, so the new model must prefill the relevant context before it emits a token. Across representative agentic workloads, including a production coding-agent trace with a mean context of 63K tokens, the prefix hit rate falls from 80-89% on same model calls to 44-65% after a switch, and a switch call takes 1.5-3.2x as long. We present SHADOWPREFILL, which prefills the session's context on the model the router has not chosen, in capacity that instances serving normal traffic leave idle, so the next switch finds its prefix cached. SHADOWPREFILL harvests the per-iteration SLO slack of requests in flight, protects routed calls, spills shadow KV to host memory, and falls back to a cold switch whenever the shadow is incomplete. With LLMRouter routing 24 model pairs over five agentic benchmark suites, SHADOWPREFILL cuts switch-call latency by 8.5-21.4%, LLM-call latency by 3.7-17.7%, and routed token cost by 22.5-46.1%, while normal traffic keeps its SLO and energy per request rises by 0.4%.

DramaGrow: A Regime-Aware Benchmark for Auditing LLM Judges of Multi-Episode Drama

arXiv:2611.09938v1 Announce Type: new Abstract: LLMs can now draft complete multi-episode dramas, and LLM judges are increasingly used to evaluate them. Assessing the reliability of these judges is difficult because drama quality encompasses specification-backed errors, reproducible preferences, and legitimate disagreement among human raters. Most existing evaluations reduce these cases to a single score from one judge call and leave tie calibration, call-to-call variation, and the vote margin underlying the human reference unreported. We introduce DramaGrow, a regime-aware benchmark and audit protocol for AI-generated Chinese five-episode drama. The benchmark assigns verifiable constraint violations, four-of-five preferences, and within-panel disagreements to separate denominators and evidence rules. It contains 800 episodes, 120 byte-preserving interventions, and 1,155 blinded five-rater records, including a frozen extension with 114 supermajority pairs and 21 disagreement pairs. On the frozen sets of 114 preference pairs and 35 constraint interventions, the repeated absolute-score rubrics yield overlapping performance, with mean accuracies of 53.8-55.8% and three-call agreement of 29.8-32.5%, and their accuracy falls from 71.6-71.9% to 34.8-37.1% as the tie band widens. DirectPair-GLM reaches 85.1% preference accuracy, 89.5% constraint F1, and 81.9% Detect+Locate on the same items, and exact subpanel enumeration identifies four-of-five as the smallest stable human vote margin. Evaluation protocol, tie calibration, judge-call variation, and panel margin therefore affect performance as much as rubric design. We will release the specification packages, scripts, interventions, anonymized human votes and rationales, frozen judge runs, and analysis code.

Policy Over Purpose: Auditing LLM Content Moderation of Reclaimed Slur Usage

arXiv:2611.10327v1 Announce Type: new Abstract: As Perspective API is deprecated, platforms are turning to large language models for content moderation, encoding policy as natural language that the model must interpret at inference time. Reclaimed slurs test this design directly. The same term can express in-group pride or target the community that reclaimed it, and misreading it silences the speech moderation is meant to protect. We audit four general-purpose LLMs and four moderation-specific systems, with Perspective API as a legacy baseline, on IYKYK, a corpus of reclaimed and derogatory usage of three slurs annotated by community members. Across the policy, decision, reasoning, disparity, and steering layers of this audit, we find that whether a system was purpose-built for moderation explains little about how it handles reclaimed language. No system reliably agrees with community judgment, and all fail most on language reclaimed within Black communities. What separates systems is instead how their policy is represented. Only systems whose policy is expressed in the prompt can be steered at all, and even then, explicit definitions change how models justify a decision more than how they make it. We argue that policy transparency determines whether moderation failures can be corrected, but that correcting them requires contextual judgment that definitions supplied at inference time cannot provide.

Decision Context: Persistent Behavioral Learning for AI Agents

arXiv:2611.08644v1 Announce Type: new Abstract: LLM agents do not learn from corrections across sessions. A user fixes the same mistake at the start of every conversation. Existing agent memory systems primarily handle facts, such as preferences, entities, and conversational history, rather than procedural knowledge. We present Forge Decision Context (forge_dc), a system that captures behavioral signals from human corrections and self-assessed tool-call failures, compiles them into imperative RULE directives, and injects them at the decision points where they apply. We evaluate forge_dc alongside four predecessor approaches on DC-Bench, a new benchmark of 71 scenarios across nine behavioral dimensions, built from real executed agent traces in a deterministic mocked-tool environment. forge_dc reaches 95% compliance, up from 55% with naive retrieval, while injecting approximately 10 times fewer tokens per query. Across the five-stage progression, the strongest systems represent behavioral memory as explicit directives and deliver it at the decision points where it applies. The results identify representation, activation, and placement as central design choices for persistent behavioral learning.