LLMs Can Recognize News Framing but Cannot Reverse It

A Task That Defeats Every Major Model

Give a large language model a news article and ask it to judge the article's "framing": Is it a positive or negative report? Who is portrayed as the agent and who as the patient? Which facts lead the story and which are buried at the end?

The model does well. Recognition accuracy ranges from 0.79 to 0.94 — close to human level.

Then give it a harder task: reverse the article's framing. Turn the positive frame negative, swap agents and patients, push highlighted facts into the background. One hard constraint: every atomic fact must be preserved. No fact may be deleted, and no new fact may be added. Only the way the facts are told may change.

The model fails. Full reversal success rate: 0.044 to 0.068.

This is the core finding of a paper published in September 2026, *Recognizing Is Not Reversing: The Asymmetry of Framing Inversion in LLMs*, by researchers from the University of Maryland and New York University. They built a benchmark called FramingInversion specifically to test LLMs' "framing manipulation" ability.

What Is News Framing?

News is not a list of facts. The same set of facts told differently produces completely different effects. That is "framing."

The paper identifies three core framing dimensions:

1. Evaluative Lexis. The same event can be described neutrally ("the government reduced the budget"), negatively ("the government slashed the critical budget"), or positively ("the government optimized budget allocation"). The facts do not change, but the evaluative words steer the reader's emotional response.

2. Agency Realization. "Police dispersed protesters" — police are the agent, protesters the patient. "Protesters were dispersed by police" — grammatically passive, but the focus shifts. "Protesters experienced a dispersal" — the agent vanishes entirely, as if the dispersal were a natural event. Three phrasings, same facts, but completely different implications about responsibility.

3. Information Salience. A report that puts a fact in the first sentence of the lead versus the seventeenth paragraph produces vastly different effects. Readers assume "the important stuff comes first" — a universal convention of news writing. Changing the order of facts changes the reader's judgment of what matters most.

The FramingInversion Benchmark

The researchers constructed 540 "framing pairs." Each pair contains two articles covering the same atomic facts with completely opposite framings. Construction was rigorous:

1. Atomic fact extraction: all specific factual claims are pulled from the original article. 2. Framing reversal: evaluative lexis, agency, and salience are inverted while the atomic facts stay fixed. 3. Factual fidelity verification: an independent model checks that the reversed text preserves all atomic facts.

Two task types:

• Recognition task: given one article, judge the framing direction (positive/negative), framing type (evaluative/agency/salience), and specific framing elements.
• Reversal task: given one article, generate a version with the completely opposite framing while preserving all atomic facts.

Recognition: Near Perfect

• Framing direction (positive vs. negative): 0.79–0.94 accuracy
• Framing type (evaluative/agency/salience): 0.85–0.92
• Specific framing elements (which words are evaluative, who is the agent, which facts are highlighted): 0.72–0.88

All tested models — GPT-4o, Claude 3.5 Sonnet, Llama 3.1 70B, Mistral Large — perform well. LLMs have framing perception: they can "see" the frame.

Reversal: Nearly Zero

• Full reversal success rate: 0.044–0.068 (4.4%–6.8%)
• Even when framing type and direction were correctly identified first: only 0.071

The models know the article is positively framed, know it uses evaluative lexis, know who the agent is — but they cannot turn that "knowing" into "manipulation."

Detailed decomposition:

• Factual fidelity (all atomic facts preserved): 0.84 — quite high
• Framing direction reversal: 0.31 — barely
• Full reversal (fidelity + direction reversal + consistent framing type): 0.044–0.068 — nearly zero

The most common failure mode: facts preserved but framing not truly reversed — a few word swaps and reordered sentences, but the overall frame stays the same. Or the inverse: framing reversed but facts lost — inconvenient facts deleted or new facts invented to support the new frame.

"Preserve all facts" and "fully reverse the frame" simultaneously is a combination almost no model reliably achieves.

Why Recognition ≠ Reversal

1. Recognition is classification; reversal is generation. Recognition only requires outputting a label. Reversal requires generating full text while satisfying multiple constraints at once (fact fidelity + framing reversal + fluency). More constraints, harder generation.

2. Recognition is passive understanding; reversal is active manipulation. Recognition means "reading" framing signals. Reversal means "operating" on framing — which requires understanding how the frame is constructed, not just what it is. Knowing how a tool works and knowing how to use it to do something else are different abilities.

3. Training data asymmetry. "Judge a text's framing" tasks vastly outnumber "reverse a text's framing" tasks in LLM training data. The former is a common NLP annotation task; the latter barely exists in natural data.

4. Framing is a systemic constraint, not a local edit. Reversing framing requires coordinated changes to evaluative lexis, agency structure, and information order. Changing one evaluative word may require retuning an entire paragraph's tone; changing the agent may require rewriting a sentence's grammar. This systemic coordination exceeds current LLMs.

What It Means

LLM "understanding" is incomplete. The paper draws a precise line: being able to recognize is not the same as being able to manipulate. A model can perfectly identify a frame's direction, type, and elements yet be unable to reverse it — its understanding is passive, not active.

Humans are similar. We can recognize a painting's style ("that's Impressionist") without being able to paint Impressionist works. Recognition and manipulation are separate abilities; the former does not imply the latter.

Direct implications for AI safety. If LLMs cannot reverse framing, they cannot easily construct framing either — meaning the bar for using LLMs for large-scale opinion manipulation is higher than assumed. Conversely, LLMs are limited at "deframing": you cannot expect a model to automatically rewrite a biased article into a neutral one. It can see the bias but cannot remove it.

Evaluation should separate "recognition" from "manipulation." Many benchmarks treat "understanding" as one thing — give the model a text, ask a few questions. But a model can be near-human at recognition while far below human at manipulation. Mixing them overestimates real capability.

My Take

What interests me most is the capability asymmetry the paper reveals — a pattern seen across LLM tasks:

• Can detect errors but not fix them (hallucination detection vs. elimination)
• Can detect bias but not remove it (bias detection vs. debiasing)
• Can recognize style but not produce it (style recognition vs. style transfer)
• Can recognize framing but not reverse it (this paper)

All point to the same conclusion: LLM understanding is passive, not active. Models can "see" patterns but not "operate" on them — unlike humans, who can typically generate and manipulate a pattern once they understand it.

The root may lie in training data: recognition tasks vastly outnumber manipulation tasks, since the internet is full of judgments and classifications but rarely demonstrates "how to reverse this judgment." Denser manipulation tasks might narrow the gap.

But the deeper possibility is that recognition and manipulation are simply different levels of ability. Recognition needs pattern matching; manipulation needs pattern construction. Perception and generation are separated in human brains too — you can recognize a face you cannot draw. LLMs may inherit the same separation.

Either way, the paper draws a precise boundary around "LLMs understand language": understanding to the point of recognition is not understanding to the point of manipulation. Next time someone says "the LLM understands X," ask: at the recognition level, or the manipulation level? The gap between them may be far deeper than we think.

---

Paper: Recognizing Is Not Reversing: The Asymmetry of Framing Inversion in LLMs

Authors: University of Maryland, New York University

Published: 2026-09-10