Papers for

machine learning evaluators

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Benchmark reveals flaws in how expert taste is measured in tests

How to Build a Benchmark That Actually Measures Taste

Abstract: Expert-aligned "taste" benchmarks should reward craft judgment, not the ability to game the test. We introduce TasteBench, a pairwise good-vs-bad benchmark across visual, brand, book, and tech items under a frozen rubric with expert-only gold (panel n>=3; percent agreement >=0.60 primary gate; weak engagement labels train/analysis-only), together with a control protocol that separates craft judgment from test-gaming. Applying the protocol to a Round-1 pilot (primary_human_expert; three biological human experts; n=250 pairs) surfaces and quantifies three validity threats: two construction artifacts that inflate the trivial-baseline layer, and a difficulty ceiling that inflates every score. (1) Position: the preferred item is slot A with probability 0.956, so an always_item_a policy scores 0.956 (fair chance 0.5). (2) Lexical leakage: a filename token cue scores 0.912. (3) Difficulty ceiling: craft domains show near-ceiling human agreement (book/brand/tech 1.0), and even under A/B randomization (eval_assignment_seed7; gold_side_A ~0.47) all five VLM cards reach 1.0 on craft, with only visual separating (~0.84-0.91; a post-hoc hard subset, n=107, leaves craft at 1.0). The protocol (expert-only gold, a weak-label firewall, side randomization at construction, filename scrubbing, and hard filters) makes each artifact measurable and, for the two construction artifacts, correctable; we release a reference eval harness and a side-randomization tool, and show the tool removes the position skew (0.916 to 0.495/0.505). The result is a calibrated view of what the benchmark does and does not yet measure: position and leakage are fixable, whereas the craft-domain ceiling is a pair-difficulty limit that Round-2 must address. We do not evaluate engagement, era/circle conditioning, or calibration here, and we claim no general taste SOTA.

Sat 10 OctArtificial IntelligenceComputation and LanguageComputer Vision and Pattern Recognition
The gist
Getting computers to judge taste, like choosing which book or brand is better, is tricky because some tests can be easily tricked. The authors created a new test called TasteBench that uses experts' opinions and special rules to make sure the test really measures good taste, not just who can cheat it. They found problems in the way previous tests were made, like always favoring one option or leaking hints that can be guessed. Their methods fix some of these problems but also show that some taste judgments are so hard that even experts fully agree, making it hard to improve further. More work is needed to make taste tests better and more reliable.
Open → 2610.14707v1

Language model self-reports do not prove the presence of self awareness

Self-Reports Do Not Identify Self-Models: An Identifiability Test for Counterfactual Reports

Abstract: Language-model self-reports are evidence about behavior in a prompt environment, not by themselves evidence of a self-model. We investigate counterfactual reports about affect-like states under activation interventions and ask whether the report remains bound to the named intervention when the demonstration environment changes. Across three open instruction models, wrong-source demonstrations move reports toward the source answer family, while explicit mechanism binding reduces this pull. Self-report benchmarks should include environment-shift invariance tests under fixed intervention before treating accuracy as evidence for an autonomous report mechanism.

Sat 26 SeptComputation and Language
The gist
This paper shows that when language models say what they would do in different situations, it doesn’t necessarily mean they have a real understanding or self-awareness. The researchers tested if the models’ statements about feelings would stick to the same cause when the setting changed, and found they often don’t. This means using these self-reports as proof that language models have a sense of self is unreliable.
Open → 2609.32449v1

Reasoning models often fail on similar rule learning tasks

Thought without systematicity? Evaluating reasoning models on rule induction tasks

Abstract: A central tenet of human cognition is systematicity, the principle that understanding one concept is inherently tied to understanding close variations of that concept. Do reasoning models robustly exhibit such systematicity? If so, we would expect consistent performance on structurally equivalent variants of the same task. Here, we extend established rule induction tasks from cognitive science to assess the systematicity of thought in current reasoning models. Each task family has compositional structure that we use to create structurally equivalent task variations through task isomorphisms such as recombination and substitution. We find that despite being able to correctly solve a task, models often fail on structurally equivalent variants of the same task. These findings suggest that many model behaviors lack systematicity, rendering it difficult to robustly establish the cognitive abilities of reasoning models beyond the particular contexts they were evaluated in.

Sat 12 SeptComputation and LanguageArtificial IntelligenceMachine Learning
The gist
The paper looks at whether reasoning models think in a way that lets them apply what they learn to slightly different but related problems. The authors tested these models on different versions of rule-learning tasks that are basically the same but with small changes. They found that even when models do well on one task, they often fail on similar tasks with the same underlying logic. This means these models might not understand or think as systematically as humans do.
Open → 2609.13948v1