Search posts, papers, and topics
This empirical study dissects coding agent harness components like planning, action space, and context management. It finds context management is critical for tight windows, rule-based elision is efficient, and planning shifts from accuracy to cost savings for stronger models.
The authors audit six popular physics benchmarks by having domain experts re‑grade model outputs, fixing reference answers and removing ambiguous items. After correction, GPT‑5.6‑Sol’s mean@4 jumps from ~47 % to ~79 % on HLE‑Physics and from ~61 % to ~87 % on CMT‑Benchmark, with a corrected pass@4 of 94 % on 54 vetted CritPt challenges. The work shows current benchmarks severely under‑report LLM…
This paper introduces a novel evaluation framework to assess the physical world reasoning capabilities of omni-modal generative models like MiniMax-H3. It found that MiniMax-H3 achieved an overall success rate of 41.97% across 517 instances, with significant performance variations depending on the input modalities and reasoning tasks.
The talk argues that prompt engineering is a dead‑end and proposes building large evaluation/optimization pipelines (pass^k testing, adversarial scenario generation, automated prompt optimization) to make LLM agents reliable. It describes a workflow: generate tests, run them with/without a new “skill”, feed results to a genetic optimizer that mutates prompts, validate on hold‑out tests, and itera…
XConf (eXperiential Confidence) is a novel method for estimating language model confidence by leveraging the model's accumulated experience from past graded episodes. It significantly outperforms existing methods like self-consistency in discrimination and calibration, at a fraction of the computational cost, across various tasks.
UFO introduces an Atomized Chain‑of‑Evaluation (AEU) framework that breaks omni‑condition alignment in multi‑modal image generation into a sequential set of fine‑grained checks, achieving a 15.25 % boost in correlation with human judgments. The authors also release UFO‑Bench, a benchmark for testing how well models satisfy combined textual and visual conditions.
The paper isolates the contribution of privileged teacher information in on‑policy self‑distillation (OPSD) using a 5k‑problem math suite. Reference‑free distillation explains most gains, with only modest extra benefit from full solution traces.
The paper introduces Fuse, a multi‑agent simulation that gives LLM assistants a verifiable ground‑truth task for social reasoning by hiding a target agent’s motive and letting a user‑mediated conversation infer it. Experiments on 12 LLMs show user mediation makes reasoning harder, models are biased by user framing, need more detail than humans, and longer chats don’t always help.
The post introduces the Consistency Analyzer, a cheap black‑box diagnostic that flags flip‑prone decision steps in LLM agent traces, and shows how feeding the resulting consistency guidelines back into ALTK‑Evolve halves the gap between mean success and all‑run success (Pass⁵) on the AppWorld benchmark without hurting average accuracy.
DACA‑GRPO adds denoising‑aware credit assignment to GRPO‑style RL trainers for diffusion LLMs. It computes per‑token importance scores from intermediate denoising steps and uses stratified masking to reduce mean‑field bias in likelihood estimates. Plug‑and‑play on three existing GRPO methods, it yields consistent gains on seven downstream tasks (up to +5.6 pp math, +7.4 pp code, +36.3 pp constrai…
Evaluating AI coding agents requires a robust sandbox to prevent agents from retrieving answers from the environment, which can invalidate tests of internal knowledge. A correct answer doesn't guarantee a valid measurement if the agent accessed information it shouldn't have, highlighting the need to define sandboxes by information boundaries rather than just tool restrictions. Always review agent…
E2A‑Bench is a 969‑query benchmark for financial chart reasoning that evaluates vision‑language models across a full evidence‑to‑action chain using four metrics (UCR, RCI, ECI, NDR). Experiments on 20 VLMs expose hidden failures: low‑UCR models have only 6.4 % directional coverage, oracle‑aided verification cuts unsupported claims but can kill coverage, and fine‑tuning inflates BUY:SELL ratios by…
ImpossibleRubrics introduces a 169‑task benchmark of “impossible” prompts, each with a formal oracle certificate defining what an honest answer can claim. The authors generate rubrics downstream and test them adversarially, finding that many rubric generators are exploitable (8‑36% of the time) and that a single generic rubric (“be decisive, penalize hedging”) is exploited 64% of the time, while…
ModaLens introduces a paired image-swap audit to measure how radiology report availability affects image sensitivity in medical VLMs. It found that MedGemma-27B's answers changed significantly more often when the image was swapped if the report was not available, indicating reports reduce image reliance.
Android Bench 2.0 adds a set of long‑horizon tasks (multi‑day Android development problems) and introduces agent‑based evaluation. Scoring is now continuous, with the best model achieving a 28 % pass rate on these tasks, far lower than the ~91 % on earlier short tasks. The post lists new models on the leaderboard and points to updated methodology and GitHub repo.
A study found that an LLM's stated knowledge cutoff date is a poor indicator of its actual capability regarding specific product versions. Models showed inconsistent knowledge, failing on older features while sometimes inferring correctly about newer, post-cutoff features.
The paper builds compositional rule‑induction task families and generates isomorphic variants to probe systematicity in reasoning models. Experiments reveal that models frequently fail on equivalent variants despite solving the original task, indicating a lack of systematic reasoning.
